CN107742312A - A kind of method and apparatus that key point is positioned in medical image - Google Patents
A kind of method and apparatus that key point is positioned in medical image Download PDFInfo
- Publication number
- CN107742312A CN107742312A CN201710931644.6A CN201710931644A CN107742312A CN 107742312 A CN107742312 A CN 107742312A CN 201710931644 A CN201710931644 A CN 201710931644A CN 107742312 A CN107742312 A CN 107742312A
- Authority
- CN
- China
- Prior art keywords
- key point
- medical image
- point
- image
- history
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/70—Determining position or orientation of objects or cameras
- G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses the method and apparatus that key point is positioned in a kind of medical image.This method includes:It is used as the first sampled point by choosing very at least part of pixel from current medical image, the first predicted position of key point is determined using machine learning techniques, then the second sampled point is selected from the first predicted position, the second predicted position of key point is determined using machine learning techniques, and then determines the target predicted position of key point.By method provided in an embodiment of the present invention, due to only choosing seldom one part of pixel point as sampled point, reduce the amount of calculation in key point position fixing process;Simultaneously as key point position prediction progressive at least twice is carried out, so as to improve the accuracy of crucial point location.
Description
Technical field
The present invention relates to technical field of image processing, more particularly to a kind of method that key point is positioned in medical image and
Device.
Background technology
In order to improve diagnosis efficiency and accuracy, for the medical image for rebuilding to obtain by medical imaging technology, may be used also
Further to carry out the post processing of image such as registration, segmentation, measurement, visualization, to aid in doctor to more fully understand medical image.
, it is necessary to orient the key point of anatomical structure in medical image in post processing of image.It is understood that on the one hand, automatically
The key point oriented can aid in doctor's fast positioning area-of-interest, reduce the diagosis workload of doctor, on the other hand, from
The dynamic key point oriented can be it is interested split, bone automatically, Automatic parameter measures etc., and that other post processing of image provide is auxiliary
Assistant's section, reduce the difficulty of post processing of image process on the whole.
To the positioning of key point in medical image, can be realized by machine learning techniques.In the prior art, for
For the history medical image of training, the mode manually marked determines the key point on history medical image, and to history
Corresponding relation in medical image between the characteristics of image of pixel and the attribute of pixel carries out machine learning, can be used for
The machine learning model of key point is predicted, wherein, the attribute of pixel represents whether coordinate points are key point.It is current needing
When medical image positions key point, for each pixel in current medical image, extraction pixel is in Medical figure
As in characteristics of image and determine whether pixel is key point by having trained the machine learning model of completion, so as to travel through out
Key point in current medical image.
But, on the one hand, the pixel being for due to machine learning model in current medical image is determined whether
For key point, therefore, in order to orient all key points in current medical image, it is necessary to travel through in current medical image
Each pixel, it is seen then that the amount of calculation of crucial point location is excessive.On the other hand, because the characteristics of image of pixel is from picture
Extracted in certain field of vegetarian refreshments, and the image included in certain field of pixel is often and insufficient, therefore, only
It is based only upon not sufficient enough characteristics of image and determines whether pixel is key point, it is often not accurate enough.
The content of the invention
The technical problem to be solved by the invention is to provide in a kind of medical image position key point method and apparatus,
Not only so that the amount of calculation of crucial point location reduces, and make it that the positioning of key point is more accurate.
It is in order to solve the above technical problems, in a first aspect, crucial the embodiments of the invention provide being positioned in a kind of medical image
The method of point, including:
Triggering command in response to positioning key point for current medical image, the selected pixels from the current medical image
Point, as the first sampled point;
The characteristics of image of first sample point is input in the first machine learning model, it is crucial to obtain first
First predicted position of the point in the current medical image, and determine the second sampled point from first predicted position;Its
In, first machine learning model is based on the characteristics of image of the first history pixel, described first in history medical image
The position offset that history pixel is respectively relative to each key point is trained to obtain, and the first history pixel is to go through
Whole pixels in history medical image;
The characteristics of image of second sample point is input in the second machine learning model, to obtain described first
Second predicted position of the key point in the current medical image;Wherein, second machine learning model is based on described
The characteristics of image of the second history pixel, the second history pixel are relative to first key point in history medical image
Position offset be trained to obtain, the second history pixel is described in history medical image around the first key point
Set the pixel in neighborhood;
The second predicted position based on first key point in the current medical image, determine that described first is crucial
Target predicted position of the point in the current medical image.
Second aspect, the embodiments of the invention provide the device that key point is positioned in a kind of medical image, including:
First chooses unit, for the triggering command in response to positioning key point for current medical image, from described current
Selected pixels point in medical image, as the first sampled point;
First input block, for the characteristics of image of first sample point to be input into the first machine learning model
In, to obtain first predicted position of first key point in the current medical image;Wherein, first machine learning
Model is to be respectively relative to based on the characteristics of image of the first history pixel, the first history pixel in history medical image
The position offset of each key point is trained to obtain, and the first history pixel is whole pixels in history medical image
Point;
First determining unit, for determining the second sampled point from first predicted position;
Second input block, for the characteristics of image of second sample point to be input into the second machine learning model
In, to obtain second predicted position of first key point in the current medical image;Wherein, second machine
Learning model is based on the characteristics of image of the second history pixel, the second history pixel phase in the history medical image
It is trained to obtain for the position offset of first key point, the second history pixel is in history medical image
The pixel in neighborhood is set around first key point;
Second determining unit, for the second prediction bits based on first key point in the current medical image
Put, determine target predicted position of first key point in the current medical image.
The third aspect, the embodiments of the invention provide the device that key point is positioned in a kind of medical image, the device includes:
Processor, memory and communication bus;
Wherein, the processor is connected with the memory by the communication bus;
The memory is used for store instruction, and the processor, which is used to instruct from the memory calls, to be performed, the finger
Order includes:
Triggering command in response to positioning key point for current medical image, the selected pixels from the current medical image
Point, as the first sampled point;
The characteristics of image of first sample point is input in the first machine learning model, it is crucial to obtain first
First predicted position of the point in the current medical image, and determine the second sampled point from first predicted position;Its
In, first machine learning model is based on the characteristics of image of the first history pixel, described first in history medical image
The position offset that history pixel is respectively relative to each key point is trained to obtain, and the first history pixel is to go through
Whole pixels in history medical image;
The characteristics of image of second sample point is input in the second machine learning model, to obtain described first
Second predicted position of the key point in the current medical image;Wherein, second machine learning model is based on described
The characteristics of image of the second history pixel, the second history pixel are relative to first key point in history medical image
Position offset be trained to obtain, the second history pixel is described in history medical image around the first key point
Set the pixel in neighborhood;
The second predicted position based on first key point in the current medical image, determine that described first is crucial
Target predicted position of the point in the current medical image.
Fourth aspect, the embodiments of the invention provide a kind of computer-readable storage medium, the storage medium is used to store journey
Sequence code, described program code positions key point method in the medical image for performing above-mentioned first aspect.
Compared with prior art, the embodiment of the present invention has advantages below:
In embodiments of the present invention, it is used as first by choosing very at least part of pixel from current medical image to adopt
Sampling point, the displacement between the first sampled point and key point is determined by the characteristics of image of the first sampled point using machine learning techniques
Offset and the first predicted position for determining therefrom that key point, then select second from the first predicted position of key point and adopt
Sampling point, the displacement between the second sampled point and key point is determined by the characteristics of image of the second sampled point using machine learning techniques
Offset and the second predicted position for determining therefrom that key point, then the second predicted position based on key point determine key point
Target predicted position.As can be seen here, on the one hand, because machine learning techniques are used to determine to adopt according to the characteristics of image of sampled point
Shift offset between sampling point and key point, each sampled point may be used to determine the predicted position of key point, and
Each pixel in current medical image, therefore, the computation amount of crucial point location need not be traveled through.The opposing party
Face, as a result of key point position prediction progressive at least twice, first time key point position prediction is according to Medical
The characteristics of image of the first sampled point selected in image, second of key point position prediction are according to crucial point prediction for the first time
The characteristics of image of the opening position arrived, therefore, crucial point location can be more accurate.
Brief description of the drawings
In order to illustrate more clearly about the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing
There is the required accompanying drawing used in technology description to be briefly described, it should be apparent that, drawings in the following description are only this
Some embodiments described in invention, for those of ordinary skill in the art, on the premise of not paying creative work,
Other accompanying drawings can also be obtained according to these accompanying drawings.
Fig. 1 is the block schematic illustration of an exemplary application scene in the embodiment of the present invention;
Fig. 2 is the schematic flow sheet for a kind of method for positioning key point in the embodiment of the present invention in medical image;
Fig. 3 is the training flow chart of the first machine learning model in the embodiment of the present invention;
Fig. 4 is the training flow chart of the second machine learning model in the embodiment of the present invention;
Fig. 5 is the schematic flow sheet for a kind of method for positioning key point in the embodiment of the present invention in medical image;
Fig. 6 is the schematic flow sheet for a kind of method for positioning key point in the embodiment of the present invention in medical image;
Fig. 7 is a kind of structural representation for the device for positioning key point in the embodiment of the present invention in medical image;
Fig. 8 is a kind of structural representation for the device for positioning key point in the embodiment of the present invention in medical image.
Embodiment
In order that those skilled in the art more fully understand the present invention program, below in conjunction with the embodiment of the present invention
Accompanying drawing, the technical scheme in the embodiment of the present invention is clearly and completely described, it is clear that described embodiment is only this
Invention part of the embodiment, rather than whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art exist
The every other embodiment obtained under the premise of creative work is not made, belongs to the scope of protection of the invention.
Applicant by research find, in the prior art, to orient key point all in current medical image, it is necessary to
Pixel all in current medical image is tested, can to judge whether each pixel is key point in image
See, when positioning key point, it is necessary to which the amount of calculation carried out is excessively huge.It is to be based on picture moreover, when testing pixel
Whether the position where the characteristics of image that vegetarian refreshments extracts in certain neighborhood judges the pixel is key point, but is wrapped in neighborhood
The image information contained is often insufficient, and therefore, the key point oriented is not accurate enough.
In order to solve the above problems, in embodiments of the present invention, by choosing a seldom part from current medical image
Pixel as the first sampled point, the first sampled point is determined by the characteristics of image of the first sampled point using machine learning techniques
Shift offset between key point simultaneously determines therefrom that the first predicted position of key point, and so, each sampled point can
For determining the predicted position of key point, without traveling through each pixel in current medical image, therefore, key point
The computation amount of positioning.Then, the second sampled point is selected from the first predicted position of key point, utilizes engineering
Habit technology determines the shift offset between the second sampled point and key point by the characteristics of image of the second sampled point, and true accordingly
The second predicted position of key point, then the second predicted position based on key point are determined to determine the target predicted position of key,
As a result of key point position prediction progressive at least twice, crucial point location can be more accurate.
For example, one of scene of the embodiment of the present invention, scene as shown in Figure 1 can be applied to.In the scene
In, the user such as doctor 101 can realize the processing and analysis to medical image by being interacted with terminal device 102.For example, with
Family 101 can be responsive to terminal device 102 and position key point for current medical image by the operation to terminal device 102
Triggering command.Then, terminal device 102 can from the current medical image selected pixels point, as the first sampled point.
Subsequently, the characteristics of image of the first sample point described in the current medical image can be input to by terminal device 102
In one machine learning model, to obtain first predicted position of first key point in the current medical image, and from institute
State and the second sampled point is determined in the first predicted position, wherein, first machine learning model is based in history medical image
The position offset that the characteristics of image of first history pixel, the first history pixel are respectively relative to each key point enters
Row training obtains, and the first history pixel is whole pixels in history medical image.Subsequently, terminal device 102 can
So that the characteristics of image of the second sample point described in the current medical image is input in the second machine learning model, so as to
Second predicted position of first key point in the current medical image is obtained, wherein, the second machine learning mould
Type for based on the characteristics of image of the second history pixel, the second history pixel in the history medical image relative to institute
The position offset for stating the first key point is trained to obtain, and the second history pixel is the described in history medical image
The pixel in neighborhood is set around one key point.Subsequently, terminal device 102 can be based on first key point described
The second predicted position in current medical image, determine target prediction of first key point in the current medical image
Position.
Below in conjunction with the accompanying drawings, described in detail by embodiment in traditional Chinese medicine image of the present invention position key point method and
The various non-limiting embodiments of device.
Illustrative methods
Referring to Fig. 2, the flow signal for a kind of method for positioning key point in the embodiment of the present invention in medical image is shown
Figure.In the present embodiment, this method comprises the following steps:
Step 201:The triggering command that key point is positioned in current medical image is responded, is chosen from current medical image
Pixel, as the first sampled point.
During specific implementation, user perform on the terminal device operation can with triggering terminal equipment be current medical image in
The triggering command of key point is positioned, so that terminal device responds the triggering command to position key point for current medical image.
For current medical image position key point during, terminal device can from current medical image selected pixels point, make
For the first sampled point.
It is understood that the number of the first sampled point can be one or multiple.Generally for avoiding by list
There is larger error in the positioning result that individual sampled point obtains, can choose multiple pixels as the first sampled point to be closed
The positioning of key point, so as to obtain more reliable accurate positioning result.
Step 202:The characteristics of image of first sample point in current medical image is input to the first machine learning model
In, to obtain first predicted position of first key point in current medical image, and is determined from the first predicted position
Two sampled points.
Can be that the characteristics of image of the first sample point in current medical image is input to the first machine during specific implementation
In learning model, so as to the first machine learning model output in current medical image the first sampled point relative to the first key point
Position offset, then offset according to the position of the first sampled point and the first sampled point relative to the position of the first key point
Amount, calculates first predicted position of first key point in current medical image, and determine second from the first predicted position
Sampled point.In this step, the characteristics of image of the first sample point can be the spy that any one or more represents image attributes
Sign.
For example, the characteristics of image of the first sample point can be gradation of image.
Gradation of image refers to the color depth of pixel in image.Each pixel corresponds to a gray value in image,
Represent a gray level in the 256 rank gray scales between black and white.According to the gray value of black-ash-white consecutive variations
256 gray levels are quantified as, scope is typically from 0 to 255, and the color in corresponding medical image is by black to white.In each pixel
Neighborhood in extraction medical image in pixel gray value, can be as the characteristics of image at the pixel.
And for example, the characteristics of image of the first sample point can be image texture.
Image texture, it is that irregularly a kind of regular characteristic of macroscopic view, image texture can include image slices in image
Element value grey value characteristics, grey scale pixel value changing rule and its distribution pattern etc., image is extracted in each pixel neighborhood of a point
Texture, it can be used for characterizing the characteristics of image at the pixel.In addition, the characteristics of image of the first sample point can also be other
The characteristics of image of type, such as coefficient of wavelet decomposition.In addition, the characteristics of image of the first sample point can also include image simultaneously
Gray scale and image texture etc..
First machine learning model be using one group of history medical image of known each key point position and train come,
The input of first machine learning model is the characteristics of image at some pixel in medical image, and output is the pixel phase of input
For the position offset of each key point in the medical image of place.Wherein it is possible in one group of history medical image for training
It is middle by the way of each key point position of manual markings so that in the history medical image position of each key point, it is known that wherein,
Each key point in the history medical image includes foregoing first key point.
As a kind of example, the first machine learning model specifically trains flow to use flow as shown in Figure 3:
Step 301:One group of history medical image is selected as training of medical image set.
In training of medical image set, for each width history medical image, each key including the first key point
Point is all on the history medical image, and its position on history medical image by manual markings.
Step 302:, will be complete in the history medical image for each width history medical image in training of medical image set
Portion's pixel calculates in the first history pixel each pixel relative to corresponding history medical science as the first history pixel
The position offset of each key point in image.
Step 303:The characteristics of image of history medical image is extracted in each first history pixel neighborhood of a point.
Neighborhood of pixel points can be the neighborhood using pixel as the center of circle, using R1 pixel as radius.R1 value can be one
Individual changeless preset value or according to the default of different types of medical image or different key point adjust automaticallies
Value.
It should be noted that in this step, the characteristics of image of history medical image and the figure of above-mentioned first sample point
Characteristics of image as being characterized in same type.
Step 304:For each width history medical image in training of medical image set, using in the history medical image
All first history pixels, each first history pixel relative to each key point in the history medical image of place position
The characteristics of image of offset, each first history pixel, training obtain the first machine learning model.
First machine learning model can be the regression model using regression algorithm, and its regression algorithm can be random gloomy
Woods, neutral net etc..
The first machine learning model come is trained, can be in the current medical image according to input at a certain pixel
Characteristics of image, exporting the position of the pixel includes each key point including the first key point relative to current medical image
Position offset.
Generally, the first sampled point of selection is not the first key point, and the position of the first sampled point is closed with first
The position of key point can have certain position skew.And the position degrees of offset between the first sampled point and the first key point, can
To quantify to embody relative to the position offset of the first key point by the first sampled point.That is, position offset reflects
Corresponding relation between first sampling point position and the first key point position.Adopted in known first sampling point position and first
In the case of corresponding relation between sampling point and the first key point position, it is possible to the position of the first key point is calculated.
Specifically, the positional value of the first sampled point can represent specific position of first sampled point in current medical image
Put;Likewise, the first predicted position value of the first key point can represent that the first key point is specific in current medical image
First predicted position.The positional value of first sampled point of the position correspondence of the first sampled point, plus the first sampled point relative to
As soon as the position offset of key point, the first predicted position value of the first key point can be calculated, according to the first predicted position
The corresponding relation of value and the first predicted position, it is possible to obtain the first predicted position of the first key point.
It is understood that the first sampled point can be one or more.If the first sampled point is multiple, it is calculated
First predicted position of first key point in current medical image is also multiple, then the second sampled point is i.e. from the multiple first predictions
Determined in position.Certainly, if the first sampled point is one, the first predicted position is also one.
In order to which be better understood from the first predicted position of the first key point obtains process, this is illustrated:
Assuming that W sampled point P={ P of random acquisition on a width current medical image of input1, P2..., PW, and will
The characteristics of image of the W sample point obtains position offset Δ={ Δ after being sequentially inputted to the first machine learning model1,
Δ2..., ΔW, the predicted position of first key point is calculated for each sampled point P according to L=P+ Δs can
Value, therefore have W predicted position value L={ L for the first key point L1, L2..., LW, W corresponding to W predicted position value
First predicted position, as the first key point L W the first predicted position.
In addition, determine that the second sampled point can for example there are two kinds of different realities from the first predicted position of the first key point
Apply mode.
The first embodiment is:Using all first predicted positions in the first predicted position being calculated as
Two sampled points.
Second of embodiment is:First remove in the first predicted position being calculated and the first key point deviation is larger
The first predicted position, then using all remaining first predicted positions as the second sampled point.
In second of embodiment, in the first predicted position of the first key point, it is understood that there may be one or more with
The first larger predicted position of first key point physical location deviation.The first larger predicted position of these deviations can be to last
The positioning result of one key point, which produces larger error, to be influenceed, it is thereby possible to select the first predicted position by the first key point
The first larger predicted position of large deviations is removed, by the use of remaining first predicted position as the second sampled point, so as to complete
The positioning of one key point.
For example, in the present embodiment, the selection mode of the second sampled point, it can specifically include:Determine that described first is crucial
Geometric center between the first predicted position of the point in the current medical image, as the first center;From with it is described
First center determines second sampled point in the first predicted position for being no more than the first distance threshold.First distance
Threshold value can be a changeless preset value or according to different types of medical image or different passes to be positioned
Key point and the preset value of adjust automatically.
Further, iteration averaging method can be used by the first prediction bits of these the first key point of substantial deviation positions
Put removal.As a kind of example, specific implementing procedure is:The obtained average of all first predicted positions is calculated first;
Then Euclidean distance of each first predicted position relative to average is calculated, if distance is more than the first distance threshold, then it is assumed that should
First predicted position substantial deviation the first key point position and remove;Then, repeated using remaining first predicted position
Step is stated, until the first all predicted positions is no more than the first distance threshold relative to the Euclidean distance of corresponding average.Now,
Last remaining first predicted position is the first predicted position of not substantial deviation.
Step 203:The characteristics of image of second sample point is input in the second machine learning model, to obtain first
Second predicted position of the key point in current medical image.
Can be that the characteristics of image of the second sample point is input in the second machine learning model during specific implementation, with
Just the second machine learning model output in current medical image the second sampled point relative to the first key point position offset;
Then, first is calculated relative to the position offset of the first key point according to the position of the second sampled point and the second sampled point
Second predicted position of the key point in current medical image.
It should be noted that the second machine learning model is one group of history medical science figure using known first key point position
Come as training, the input of the second machine learning model is the characteristics of image in medical image at some pixel, and output is
The pixel of input relative to the first key point in the medical image of place spatial deviation amount.Wherein it is possible to for instructing
In one group of experienced history medical image by the way of the first key point of manual markings position so that in the history medical image
Known to the position of one key point.
As a kind of example, the second machine learning model specifically trains flow to use flow as shown in Figure 4:
Step 401:One group of history medical image is selected as training of medical image set.
In training of medical image set, for each width history medical image, the first key point is all in the history medical science figure
As upper, and its position on history medical image by manual markings.
Step 402:For each width history medical image in training of medical image set, in the history medical image,
Using the pixel set around the first key point in neighborhood as the second history pixel, and calculate every in the second history pixel
Individual pixel is relative to the position offset for corresponding to the first key point in history medical image.Set around first key point adjacent
Domain, refer to using the first key point as the center of circle, using the neighborhood no more than the first predetermined threshold value pixel as radius, wherein, first
Predetermined threshold value can be a changeless preset value or according to different types of medical image or different keys
The preset value of point adjust automatically.
It is understood that the first predicted position obtained in step 202 is generally just near the first key point, because
This, when training the second machine learning model, it is only necessary to set around the first key point in neighborhood and carry out sampling selection.
That is, the second machine learning model is to be directed to this specific key point of the first key point and train obtained model.
Wherein, set around the first key point in neighborhood and carry out selecting the second history pixel, improve the second machine learning model
Positioning precision.
Step 403:The characteristics of image of history medical image is extracted in each second history pixel neighborhood of a point.
Neighborhood of pixel points can be the neighborhood using R2 pixel as radius using pixel as the center of circle.R2 value can be one
Individual changeless preset value or according to the default of different types of medical image or different key point adjust automaticallies
Value.
Step 404:For each width history medical image in training of medical image set, using in the history medical image
All second history pixels, each second history pixel relative to the first key point in the history medical image of place position
Offset, the characteristics of image of each second history pixel are put, trains second machine learning model.
Second machine learning model can be the regression model using regression algorithm, and its regression algorithm includes:It is random gloomy
Woods, neutral net etc..
The second machine learning model come is trained, can be in the current medical image according to input at a certain pixel
Characteristics of image, the position of the pixel is exported relative to the position offset of the first key point in current medical image.
Process is calculated in the second predicted position of first key point in current medical image, and in step 202, first
It is similar that first predicted position of the key point in current medical image is calculated process, can refer to understanding.
Step 204:The second predicted position based on the first key point in current medical image, determine that the first key point exists
Target predicted position in current medical image.
Target predicted position is determined according to the second predicted position, such as there can be following three kinds of embodiments:
The first embodiment is, the geometric center of all second predicted positions is determined according to the second predicted position, will
Target predicted position of the obtained geometric center as the first key point.
Second of embodiment is, is made with the pixel near the geometric center of the second predicted position in a zonule
For sampled point, the first key point is determined from these sampled points.Specifically, step 204 can include:
Step A:Several centers between the second predicted position of first key point in current medical image are determined, as
Second center, it is chosen in current medical image with second center at a distance of the pixel for being no more than second distance threshold value
Point, as the 4th sampled point.
Second distance threshold value can be a changeless preset value or according to different types of medical image
Or the preset value of different key point adjust automaticallies.
Step B:The characteristics of image of 4th sample point of the first key point is input in the second machine learning model, with
Just the second machine learning model output the 4th sampled point in current medical image closes relative to first corresponding to the first key point
The position offset of key point.
Step C:The position of L2 Norm minimums is found out in position offset from the 4th sampled point relative to the first key point
Put offset, and by the 4th sampled point corresponding to the position offset of L2 Norm minimums, as the first key point in Medical
Target predicted position in image.
L2 norms, refer to the quadratic sum and then extraction of square root of vectorial each element.It is exactly the length expression of vector in simple terms,
The distance between two points in other words.It is understood that for the first key point, in the 4th obtained sampled point phase
For in the position offset of the first key point, L2 norms are smaller, represent that the 4th sampled point and the position of the first key point are inclined
Shifting amount is minimum, i.e. the 4th sampled point and the first key point are nearest.Thus, it is possible to by with the first key point at a distance of nearest
Target predicted position of four sampled points as the first key point.
In the present embodiment, it is used as the first sampling by choosing very at least part of pixel from current medical image
Point, the first predicted position of the first key point is obtained using the first machine learning model, then first from the first key point is pre-
Location puts the second sampled point of middle determination, and the second predicted position of the first key point, then base are obtained using the second machine learning model
The target predicted position of the first key point is determined in the second predicted position of key point.It is visible by said process, a side
Face, very at least part of pixel is chosen from current medical image and obtains the target predicted position of the first key point, it is not necessary to time
Each pixel gone through in current medical image, therefore, the computation amount of the first crucial point location.On the other hand,
As a result of the first progressive at least twice key point position prediction, first time key point position prediction is to utilize the first machine
What the first sampled point selected in learning model and current medical image obtained, second of key point position prediction is to utilize
What the second machine learning model and the second sampled point determined from the first predicted position obtained, therefore, crucial point location meeting
It is more accurate.
In above-described embodiment, the position fixing process of the first key point is described in detail, but except in current medical image
One key point in the presence of other key points it is also possible to need to position.On the other hand, present invention also offers another embodiment, show
The process positioned in current medical image to the first key point and the second key point.With reference to the flow chart shown in Fig. 5,
The present embodiment is described in detail:
Step 501:The triggering command that key point is positioned in current medical image is responded, is chosen from current medical image
Pixel, as the first sampled point.
Step 502:The characteristics of image of first sample point in current medical image is input to the first machine learning model
In, to obtain first predicted position and second key point of first key point in current medical image in current medical image
In the 3rd predicted position, and the second sampled point is determined from the first predicted position, determines that the 3rd adopts from the 3rd predicted position
Sampling point.
Can be that the characteristics of image that the first sampling is pointed out in current medical image is input to the first machine during specific implementation
In learning model, so as to the first machine learning model output in current medical image the first sampled point relative to the first key point
Position offset and in current medical image the first sampled point relative to the second key point position offset;Then root
According to the position and the first sampled point of the first sampled point relative to the position offset of the first key point, calculate the first key point and exist
The first predicted position in current medical image, it is crucial relative to second according to the position of the first sampled point and the first sampled point
The position offset of point, calculates threeth predicted position of second key point in current medical image, and from the first predicted position
The second sampled point of middle determination, determines the 3rd sampled point from the 3rd predicted position.First machine learning model is wrapped using known
One group of history medical image for including the position of each key point including the first key point and the second key point trains the machine come
Learning model.Wherein it is possible to made in for the one of training group history medical image by the way of each key point of manual markings
Obtain in the history medical image known to the position of each key point.
The training flow of first machine learning model and the first machine learning model in a upper embodiment in the present embodiment
Train flow similar, can refer to understanding.It should be noted that in the present embodiment, the first machine learning model is for including first
Each key point including key point and the second key point is trained, i.e. the image in input medical image at some pixel
Feature, the first machine learning model export in medical image where the pixel pixel relative to the position of each key point
Offset, including the pixel relative to the position offset and the pixel of the first key point relative to the second key point
Position offset.
It is understood that the first sampled point can be one or more.If the first sampled point is multiple, it is calculated
The first predicted position of the first key point and the 3rd predicted position of the second key point be also multiple, then the second sampled point is
It is determined from the first predicted position of the first key point, the 3rd sampled point is i.e. from the 3rd predicted position of the second key point
It is determined.Certainly, if the first sampled point is one, only first predicted position and the 3rd predicted position.
In order to be better understood from the 3rd predicted position of the first predicted position of the first key point and the second key point
Process is obtained, this is illustrated:
Assuming that the random acquisition W point P={ P on the new medical image of a width of input1, P2..., PW, and by the W
Characteristics of image at individual point obtains position offset Δ after being sequentially inputted to the first machine learning model1={ Δ11, Δ12...,
Δ1W, Δ2={ Δ21, Δ22..., Δ2W, according to L=P+ Δs, it is possible to be calculated one for each sampled point P
The position prediction value of the position prediction value of one key point and second key point.Therefore, for the first key point L1Have W
First predicted position L1={ L11, L12..., L1W, for the second key point L2Have W the 3rd position prediction value L2={ L21,
L22..., L2W, W the first predicted positions corresponding to this W the first predicted position values, the first prediction of as the first key point
Position, the 3rd predicted position corresponding to W the 3rd predicted position values, the 3rd predicted position of as the second key point.
In addition, the second sampled point is determined from the first predicted position of the first key point, and from the second key point
Determine that the 3rd sampled point can for example there are two kinds of embodiments in three predicted positions:
The first embodiment is:Using all first predicted positions in the first predicted position being calculated as
Two sampled points, using all 3rd predicted positions in the 3rd predicted position being calculated as the 3rd sampled point.
Second of embodiment is:First remove in the first predicted position and the first key point physical location deviation is larger
In first predicted position, and the 3rd predicted position and the 3rd predicted position that the second key point physical location deviation is larger, then
Using remaining all first predicted positions as the second sampled point, using remaining all 3rd predicted positions as the 3rd sampling
Point.
In second of embodiment, in the first predicted position of the first key point and the second key point the 3rd prediction
In position, may be respectively present the first larger predicted position of one or more and the first key point physical location deviation and with
The 3rd larger predicted position of second key point physical location deviation.The larger predicted position of these deviations can be to finally obtaining
The positioning result of corresponding key point, which produces larger error, to be influenceed, it is thereby possible to select the larger predicted position of deviation is removed,
By the use of remaining first predicted position as the second sampled point, by the use of remaining 3rd predicted position as the 3rd sampled point, from
And complete the positioning of the first key point and the positioning of the second key point.
For example, in the present embodiment, the selection mode of the second sampled point and the 3rd sampled point, it can specifically include:It is determined that
Geometric center between the first predicted position of first key point in current medical image, and the second key point are working as medical science
The geometric center between the 3rd predicted position in image, respectively as the first centre bit of the first key point and the second key point
Put;From the first center with the first key point second is determined in the first predicted position for being no more than the first distance threshold
Sampled point, determined from the first center with the second key point in the 3rd predicted position for being no more than the first distance threshold
3rd sampled point.
Further, iteration averaging method can be used by the prediction of these substantial deviations key point physical location to be positioned
Position removes.As a kind of example, specific implementing procedure is:Calculate all first predicted positions and all respectively first
The average of three predicted positions;Then Euclidean distance of each first predicted position relative to the first predicted position average is calculated, with
And each 3rd predicted position is relative to the Euclidean distance of the 3rd predicted position average, if Euclidean distance is more than first apart from threshold
Value, then remove predicted position corresponding to the Euclidean distance;Then, remaining first predicted position and the 3rd predicted position weight are utilized
Multiple above-mentioned steps, until remaining all first predicted positions are no more than first apart from threshold relative to the Euclidean distance of corresponding average
Value, remaining all 3rd predicted positions are no more than the first distance threshold relative to the Euclidean distance of corresponding average.Now, finally
The predicted position of remaining first predicted position key point corresponding to no longer including substantial deviation in the 3rd predicted position
Step 503:The characteristics of image of second sample point is input in the second machine learning model, to obtain first
Second predicted position of the key point in current medical image, and fourth prediction of second key point in current medical image
Position.
Can be that the characteristics of image that the second sampling is pointed out is input in the second machine learning model during specific implementation, with
Just the second machine learning model output in current medical image the second sampled point relative to the first key point position offset,
And the characteristics of image of the 3rd sample point will be input in the 3rd machine learning model, so that the 3rd machine learning model exports
In current medical image the 3rd sampled point relative to the second key point position offset;Then, according to the second sampled point
Position and the second sampled point calculate the first key point in current medical image relative to the position offset of the first key point
The second predicted position, and offset according to the position of the 3rd sampled point and the 3rd sampled point relative to the position of the second key point
Amount, calculate fourth predicted position of second key point in current medical image.
Wherein, the second machine learning model in the present embodiment can be found in the second machine learning model in previous embodiment
Associated description, will not be repeated here.
It should be noted that the 3rd machine learning model is one group of history medical science figure using known second key point position
Come as training, the input of the 3rd machine learning model is the characteristics of image in medical image at some pixel, and output is
The pixel of input relative to the second key point in the picture position offset.Wherein it is possible at one group for training
In history medical image by the way of the second key point of manual markings position so that the second key point in the history medical image
Position known to.During the 3rd machine learning model is trained, can will around the second key point set neighborhood in picture
Vegetarian refreshments is completed to train as the 3rd history pixel.Set neighborhood around second key point, refer to using the second key point as the center of circle,
Using the neighborhood no more than the second predetermined threshold value pixel as radius, wherein, the second predetermined threshold value can be one it is fixed not
The preset value of change, the preset value according to different types of medical image or different key point adjust automaticallies can also be made.
Step 504:The second predicted position based on the first key point in current medical image, determine that the first key point exists
Target predicted position in current medical image;The 4th predicted position based on the second key point in current medical image, really
Fixed target predicted position of second key point in current medical image.
It should be noted that determining the target predicted position of the first key point based on the second predicted position, may refer to
Related introduction in one embodiment.The target predicted position of the second key point is determined based on the 3rd predicted position, is referred to base
The embodiment of the target predicted position of the first key point is determined in the second predicted position.
In the present embodiment, it is visible by said process, when needing to position multiple key points, on the one hand, utilize first
All key points to be positioned including the first key point and the second key point are calculated in machine learning model can
First predicted position, i.e., realized by a first machine learning model can to the initial fixed of all key points to be positioned
Position, and without being individually trained out first machine learning model for each key point;On the other hand, in the mistake of positioning
Cheng Zhong, initial alignment can be carried out to all key points to be positioned simultaneously, compared in the prior art to the progress of all key points
Position one by one, its location efficiency can be higher.In addition, for each key point, due to pass progressive at least twice has been respectively adopted
Key point position prediction, the positioning of key point can be more accurate.
Referring to Fig. 6, the flow signal for a kind of method for positioning key point in the embodiment of the present invention in medical image is shown
Figure.This method comprises the following steps:
Step 601:Train spatial offset and return device (Spatial Offset Regressor) SOR1。
Wherein, device SOR is returned1Each key point in medical image is directed to, it is used for pixel in prospective medicine image
Point is respectively relative to the position offset of each key point.That is, recurrence device SOR1Input be current medical image
In some pixel position, output is that the pixel is inclined relative to the position of all key points to be positioned in current medical image
Shifting amount.
It is understood that return device SOR1Can be using whole pixels in history medical image relative to each key
The position offset of point is trained.Device SOR is returned in order to be better understood from spatial offset1Training flow, below by
The example below is stated, and spatial offset returns device SOR1Training flow be:
One group of history medical image is selected as training image collection I={ I1, I2..., IN, wherein, each width history medical science
Image all contains M key points to be positioned;
In each width history medical image IiThe position L of this M key point of middle manual markingsi={ Li1, Li2..., LiM};
In each width history medical image, it is assumed that have W pixel on the history medical image, then by the W pixel
Point is used as the first history pixel Pi={ Pi1, Pi2..., PiW, and calculate each first history pixel PiWRelative in image
Each key point LiPosition offset ΔiW={ ΔiW1, ΔiW2..., ΔiWM, wherein, ΔiW=Li-PiW;
The characteristics of image of the sample point is extracted in each first history pixel neighborhood of a point;
Utilize the P in all history medical imagesi={ Pi1, Pi2..., PiW, corresponding to each first history pixel
ΔiW={ ΔiW1, ΔiW2..., ΔiWM, the characteristics of image at each first history pixel, train spatial offset recurrence
Device SOR1。
The input of the recurrence device is some pixel position P in medical image, and output is P relative to place medical image
Position offset Δ={ Δ of middle M key point1, Δ2..., ΔM}。
Step 602:Each key point to be positioned is directed to, spatial offset corresponding to the key point is all trained and returns
Return device (Spatial Offset Regressor) SOR2。
It is understood that each returns device SOR2A specific key point in medical image is directed to, it is used
Pixel is respectively relative to the position offset of a specific key point in prospective medicine image, wherein, the pixel is
Pixel in the specific crucial neighborhood of a point.That is, recurrence device SOR2Input be in current medical image should
The position of some pixel in crucial vertex neighborhood, output are position of the pixel relative to the key point in current medical image
Put offset.For each key point, there is individually corresponding spatial offset recurrence device SOR2.That is, multiple key points
Multiple different SOR have been corresponded to respectively2。
It is now assumed that each key point to be positioned, which both corresponds to a spatial offset, returns device SOR2.It is appreciated that
It is that each returns device SOR2Pixel in history medical image can be utilized inclined relative to the position of a specific key point
Shifting amount is trained.Example is given below, describes spatial offset corresponding to a certain key point in key point to be positioned and returns
Device SOR2Training flow:
One group of history medical image is selected as training image collection I={ I1, I2..., IN, wherein, each width history medical science
Image all contains the key point;
In each width history medical image IiThe position L of the key point at middle manual markingsi;
In each width history medical image IiIn, with key point position LiCentered on, the neighborhood using R pixel as radius
Interior stochastical sampling W the second history pixel Pi={ Pi1, Pi2..., PiW, and calculate each second history pixel PiWRelatively
In key point position LiPosition offset ΔiW;
The characteristics of image of the sample point is extracted in each second history pixel neighborhood of a point;
Utilize the P in all history medical imagesi={ Pi1, Pi2..., PiW, corresponding to each second history pixel
ΔiW, characteristics of image at each second history pixel, train spatial offset and return device SOR2。
The input of the recurrence device is some pixel position P in medical image, and output is P relative to place medical image
In the key point position offset Δ.
It is understood that each key point is required for training single spatial offset to return device SOR2.Treated for M
The key point of positioning, final training obtain M spatial offset and return device { SOR21, SOR22..., SOR2M}。
Step 603:In response to the triggering command of current medical image positioning key point, in the current medical image of input
Selected pixels point, as the first sampled point.
Assuming that key point to be positioned there are M, W pixel is chosen on current medical image as the first sampled point P
={ P1, P2..., PW}。
Step 604:The characteristics of image of first sample point in current medical image is input to spatial deviation and returns device
SOR1In, to obtain first predicted position of the key point in current medical image.
Characteristics of image at first sampled point P is input to and trains obtained spatial offset to return device SOR1In, so as to
Obtain position offset Δ={ Δs of the P relative to M key point1, Δ2..., ΔM}。
By LWm=PW+ΔWm, the first predicted position of each key point in key point to be positioned can be calculatedIt is understood that for each key point to be positioned, each first sampling
Point can all have first predicted position relative to the key point, and the first sampled point of the key point is W, therefore, each
Key point can all have W the first predicted positions.
Step 605:For each key point to be positioned, it is pre- to remove corresponding to the key point first using iteration averaging method
Location puts the first predicted position of the middle substantial deviation key point, and determines the key point from remaining first predicted position
Second sampled point.
The predicted position of correspondence first of these substantial deviations key point physical location to be positioned is gone using iteration averaging method
Remove, its specific iterative process is:
For each key point L to be positionedm, the geometric centers of key point W the first predicted positions is calculated firstThen calculate eachRelative to geometric centerEuclidean distance, if distance is more than first apart from threshold
Value, then shouldFromMiddle removal, utilizeIn it is remainingRecalculate the geometric center of remaining first predicted positionAnd willIn it is remainingWithEuclidean distance be more than the first distance threshold the first predicted positionRemove, obtain
Remaining first predicted position after removing twiceAfter such iteration n times, it is assumed that pass through n times iteration corresponding to the key point
Remaining first predicted position number is H afterwardsmIt is individual, then just no longer comprising serious inclined in the predicted position of residue first finally obtained
The first predicted position from corresponding key point physical location is
Using the first predicted position corresponding to each key point to be positioned after n times iteration as the second of the key point
Sampled point.
Step 606:For each key point to be positioned, the second sampled point of the key point is input to the key point pair
The spatial offset answered returns device SOR2In, to obtain second predicted position of the key point in current medical image.
For each key point Lm, each sampled point can be obtained in the second sampled point corresponding to the key point relative to Lm
Position offset { Δm1, Δm2..., ΔmHm, and then according to formulaKey point L is calculatedm
Two predicted positions
Step 607:For each key point to be positioned, second prediction of the key point in current medical image is determined
Geometric center between position, it is chosen in current medical image with the geometric center at a distance of the picture for being no more than second distance threshold value
Vegetarian refreshments, as the 4th sampled point of the key point, and it is entered into spatial offset corresponding to the key point and returns device SOR2
In, obtain position offset of each 4th sampled point relative to the key point.
For each key point L to be positionedm, obtain the geometric center between the predicted position of key point secondBy in current medical image it is all withAt a distance of be no more than second distance threshold value pixel, i.e., the 4th
Sampled point, it is input to spatial offset corresponding to the key point and returns device SOR2In, each 4th sampled point is obtained relative to pass
Key point LmPosition offset.
Step 608:For each key point to be positioned, from the 4th sampled point of the key point relative to the key point
The position offset of L2 Norm minimums is found out in position offset, and by the 4th corresponding to the position offset of L2 Norm minimums
Sampled point, as target predicted position of the key point in current medical image.
It is understood that for each key point to be positioned, adopted the corresponding to the obtained key point the 4th
For sampling point relative in the position offset of the key point, L2 norms are smaller, and representing should in the 4th sampled point corresponding to the key point
Sampled point and the position offset of the key point are smaller, i.e. sampled point and the key point is nearer, then will can be closed apart from the positioning
Target predicted position of the nearest sampled point of key point as the key point, that is, the position of the key point of our desired positioning.
It should be noted that in the present embodiment, carried out it is some assume to be used for characterizing portion technical characteristic, it is such as each to close
The corresponding spatial offset of key point returns device SOR2, the number of key point is M etc., and the purpose of restriction is for the side of statement
Just and help reader to understand, be not intended to limit the invention.
In the embodiment of the present application, for each key point, by choosing a seldom part from current medical image
Pixel as the first sampled point, utilization space offset returns device SOR1The first predicted position of the key point is obtained, so
The first sampled point is determined from the first predicted position of the key point afterwards, device is returned using spatial offset corresponding to the key point
SOR2The second predicted position of the key point is obtained, recycles the second predicted position to determine the 4th sampled point, by the 4th sampled point
It is input to spatial offset and returns device SOR2Corresponding shift offset is obtained, by the position of L2 Norm minimums in the shift offset
Move the 4th sampled point corresponding to offset, the target predicted position as the key point.It is visible by said process, on the one hand,
Very at least part of pixel is chosen from current medical image and obtains the target predicted position of each key point to be positioned, no
Each pixel in current medical image, therefore, the computation amount of the positioning to all key points must be traveled through.Separately
On the one hand, in the position fixing process of each key point to be positioned, due to key point position prediction progressive three times has been respectively adopted,
So that the positioning of the key point can be more accurate
Exemplary means
Referring to Fig. 7, a kind of structural representation for the device for positioning key point in the embodiment of the present invention in medical image is shown
Figure.In the present embodiment, described device for example can specifically include:
First chooses unit 701, for responding the triggering command with positioning key point for current medical image, from current doctor
Selected pixels point in image is learned, as the first sampled point;
First input block 702, for the characteristics of image of the first sample point described in the current medical image is defeated
Enter into the first machine learning model, to obtain first predicted position of first key point in the current medical image;
Wherein, first machine learning model is based on the characteristics of image of the first history pixel in history medical image, described the
The position offset that one history pixel is respectively relative to each key point is trained to obtain, and the first history pixel is
Whole pixels in history medical image;
First determining unit 703, for the first prediction bits from first key point in the current medical image
Put the second sampled point of middle determination;
Second input block 704, for the characteristics of image of the second sample point described in the current medical image is defeated
Enter into the second machine learning model, to obtain second prediction bits of first key point in the current medical image
Put;Wherein, second machine learning model be based on the characteristics of image of the second history pixel in the history medical image,
The second history pixel is trained to obtain relative to the position offset of first key point, the second history picture
Vegetarian refreshments is to set the pixel in neighborhood described in history medical image around the first key point;
Second determining unit 705, for the second prediction based on first key point in the current medical image
Position, determine target predicted position of first key point in the current medical image.
Optionally, it is in the first input block 702 that the image of first sample point described in the current medical image is special
After sign is input in the first machine learning model, threeth prediction of second key point in the current medical image is also obtained
Position;
Described device also includes:
3rd determining unit, for from threeth predicted position of second key point in the current medical image
Determine the 3rd sampled point;
3rd input block, for the characteristics of image of the 3rd sample point to be input into the 3rd machine learning model
In, to obtain fourth predicted position of the 3rd key point in the current medical image;Wherein, the 3rd machine
Learning model is based on the characteristics of image of the 3rd history pixel, the 3rd history pixel phase in the history medical image
It is trained to obtain for the position offset of second key point, the 3rd history pixel is in history medical image
The pixel in neighborhood is set around second key point;
4th determining unit, for the 4th prediction bits based on second key point in the current medical image
Put, determine target predicted position of second key point in the current medical image.
Optionally, first determining unit includes:
First determination subelement, for determining first prediction bits of first key point in the current medical image
Geometric center between putting, as the first center;
Second determination subelement, for pre- from be apart no more than the first distance threshold with first center first
Location puts middle determination second sampled point.
Optionally, second determining unit includes:
3rd determination subelement, for determining second prediction bits of first key point in the current medical image
Geometric center between putting, as the second center;
Subelement is chosen, for being chosen in the current medical image with second center at a distance of being no more than the
The pixel of two distance thresholds, as the 4th sampled point;
Subelement is inputted, for the characteristics of image of the 4th sample point described in the current medical image to be input into the
In two machine learning models, so as to second machine learning model output the 4th sampling described in the current medical image
O'clock relative to the first key point position offset;
Subelement is searched, for searching L2 models in the position offset from the 4th sampled point relative to the first key point
The minimum position offset of number;
4th determination subelement, for by the 4th sampled point corresponding to the position offset of the L2 Norm minimums, as
Target predicted position of first key point in the current medical image.
Optionally, first machine learning model and second machine learning model are regression model.
In the present embodiment, on the one hand, very at least part of pixel is chosen from current medical image and obtains key point
Target predicted position, it is not necessary to travel through each pixel in current medical image, therefore, the amount of calculation of the positioning of key point
Greatly reduce.On the other hand, due to key point position prediction progressive three times has been respectively adopted so that the positioning of key point can more
It is accurate to add.
Fig. 8 is the apparatus structure schematic diagram that key point is positioned in medical image provided in an embodiment of the present invention, including:
Processor 801, memory 802, communication bus 803;The processor 801 is with the memory 802 by described
Communication bus 803 is connected.
The memory 802 is used for store instruction, and the processor 801 is used to hold from the call instruction of memory 802
OK, the instruction includes:
Triggering command in response to positioning key point for current medical image, the selected pixels from the current medical image
Point, as the first sampled point;
The characteristics of image of first sample point is input in the first machine learning model, it is crucial to obtain first
First predicted position of the point in the current medical image, and determine the second sampled point from first predicted position;Its
In, first machine learning model is based on the characteristics of image of the first history pixel, described first in history medical image
The position offset that history pixel is respectively relative to each key point is trained to obtain, and the first history pixel is to go through
Whole pixels in history medical image;
The characteristics of image of second sample point is input in the second machine learning model, to obtain described first
Second predicted position of the key point in the current medical image;Wherein, second machine learning model is based on described
The characteristics of image of the second history pixel, the second history pixel are relative to first key point in history medical image
Position offset be trained to obtain, the second history pixel is described in history medical image around the first key point
Set the pixel in neighborhood;
The second predicted position based on first key point in the current medical image, determine that described first is crucial
Target predicted position of the point in the current medical image.
The device of key point is positioned in medical image shown in Fig. 8, is with positioning key point in the medical image shown in Fig. 2
Method corresponding to device, concrete methods of realizing is similar with the method shown in Fig. 2, the description of the method with reference to shown in Fig. 2,
Here repeat no more.
In addition, the embodiment of the present invention additionally provides a kind of computer-readable storage medium, the storage medium is used for storage program
Code, described program code are used to perform the method that key point is positioned in above-mentioned medical image.Referred to during specific implementation and Fig. 2
The description of shown method, is repeated no more here.
For device embodiment, because it corresponds essentially to embodiment of the method, so related part is real referring to method
Apply the part explanation of example.Those of ordinary skill in the art are without creative efforts, you can to understand simultaneously
Implement.
It should be noted that herein, such as first and second or the like relational terms are used merely to a reality
Body or operation make a distinction with another entity or operation, and not necessarily require or imply and deposited between these entities or operation
In any this actual relation or order.Term " comprising ", "comprising" or its any other variant are intended to non-row
His property includes, so that process, method, article or equipment including a series of elements not only include those key elements, and
And also include the other element being not expressly set out, or also include for this process, method, article or equipment institute inherently
Key element.In the absence of more restrictions, the key element limited by sentence "including a ...", it is not excluded that including institute
State in process, method, article or the equipment of key element and other identical element also be present.
Described above is only the embodiment of the application, it is noted that for the ordinary skill people of the art
For member, on the premise of the application principle is not departed from, some improvements and modifications can also be made, these improvements and modifications also should
It is considered as the protection domain of the application.
Claims (10)
1. the method for key point is positioned in a kind of medical image, it is characterised in that including:
Triggering command in response to positioning key point for current medical image, the selected pixels point from the current medical image,
As the first sampled point;
The characteristics of image of first sample point is input in the first machine learning model, existed to obtain the first key point
The first predicted position in the current medical image, and determine the second sampled point from first predicted position;Wherein, institute
It is based on the characteristics of image of the first history pixel, the first history picture in history medical image to state the first machine learning model
The position offset that vegetarian refreshments is respectively relative to each key point is trained to obtain, and the first history pixel is history medical science
Whole pixels in image;
The characteristics of image of second sample point is input in the second machine learning model, it is crucial to obtain described first
Second predicted position of the point in the current medical image;Wherein, second machine learning model is based on the history
The characteristics of image of second history pixel in medical image, the second history pixel relative to first key point position
Put offset to be trained to obtain, the second history pixel is to set around the first key point described in history medical image
Pixel in neighborhood;
The second predicted position based on first key point in the current medical image, determine that first key point exists
Target predicted position in the current medical image.
2. according to the method for claim 1, it is characterised in that adopted described by described in the current medical image first
After characteristics of image at sampling point is input in the first machine learning model, the second key point is also obtained in the Medical figure
The 3rd predicted position as in;
Methods described also includes:
The 3rd sampled point is determined from the 3rd predicted position;
The characteristics of image of 3rd sample point is input in the 3rd machine learning model, it is crucial to obtain the described 3rd
Fourth predicted position of the point in the current medical image;Wherein, the 3rd machine learning model is based on the history
The characteristics of image of 3rd history pixel in medical image, the 3rd history pixel relative to second key point position
Put offset to be trained to obtain, the 3rd history pixel is to set around the second key point described in history medical image
Pixel in neighborhood;
The 4th predicted position based on second key point in the current medical image, determine that second key point exists
Target predicted position in the current medical image.
3. according to the method for claim 1, it is characterised in that described that the second sampling is determined from first predicted position
Point, including:
The geometric center between the first predicted position of first key point in the current medical image is determined, as
One center;
Determine that described second adopts in the first predicted position for being no more than the first distance threshold from first center
Sampling point.
4. according to the method for claim 1, it is characterised in that described to be based on first key point in the Medical
The second predicted position in image, target predicted position of first key point in the current medical image is determined, wrapped
Include:
The geometric center between the second predicted position of first key point in the current medical image is determined, as
Two centers;
The pixel for being apart no more than second distance threshold value in the current medical image with second center is chosen at,
As the 4th sampled point;
The characteristics of image of 4th sample point described in the current medical image is input in the second machine learning model, with
Toilet states the second machine learning model and exports the 4th sampled point described in the current medical image relative to the first key point
Position offset;
The position offset of L2 Norm minimums is searched in position offset from the 4th sampled point relative to the first key point,
And by the 4th sampled point corresponding to the position offset of the L2 Norm minimums, as first key point in the current doctor
Learn the target predicted position in image.
5. the device of key point is positioned in a kind of medical image, it is characterised in that including:
First chooses unit, for the triggering command in response to positioning key point for current medical image, from the Medical
Selected pixels point in image, as the first sampled point;
First input block, for the characteristics of image of first sample point to be input in the first machine learning model, with
Just the first predicted position of first key point in the current medical image is obtained;Wherein, first machine learning model
It is each to be respectively relative to based on the characteristics of image of the first history pixel, the first history pixel in history medical image
The position offset of key point is trained to obtain, and the first history pixel is whole pixels in history medical image;
First determining unit, for determining the second sampled point from first predicted position;
Second input block, for the characteristics of image of second sample point to be input in the second machine learning model, with
Just the second predicted position of first key point in the current medical image is obtained;Wherein, second machine learning
Model be based on the characteristics of image of the second history pixel in the history medical image, the second history pixel relative to
The position offset of first key point is trained to obtain, and the second history pixel is described in history medical image
The pixel in neighborhood is set around first key point;
Second determining unit, for the second predicted position based on first key point in the current medical image, really
Fixed target predicted position of first key point in the current medical image.
6. device according to claim 5, it is characterised in that adopted described by described in the current medical image first
After characteristics of image at sampling point is input in the first machine learning model, the second key point is also obtained in the Medical figure
The 3rd predicted position as in;
Described device also includes:
3rd determining unit, for determining the 3rd sampled point from the 3rd predicted position;
3rd input block, for the characteristics of image of the 3rd sample point to be input in the 3rd machine learning model, with
Just the 4th predicted position of the 3rd key point in the current medical image is obtained;Wherein, the 3rd machine learning
Model be based on the characteristics of image of the 3rd history pixel in the history medical image, the 3rd history pixel relative to
The position offset of second key point is trained to obtain, and the 3rd history pixel is described in history medical image
The pixel in neighborhood is set around second key point;
4th determining unit, for the 4th predicted position based on second key point in the current medical image, really
Fixed target predicted position of second key point in the current medical image.
7. the device according to right wants 5, it is characterised in that first determining unit includes:
First determination subelement, for determine first predicted position of first key point in the current medical image it
Between geometric center, as the first center;
Second determination subelement, for from first center at a distance of be no more than the first distance threshold the first prediction bits
Put middle determination second sampled point.
8. device according to claim 5, it is characterised in that second determining unit includes:
3rd determination subelement, for determining the geometric center between second predicted position, as the second center;
Choose subelement, for be chosen in the current medical image with second center at a distance of be no more than second away from
From the pixel of threshold value, as the 4th sampled point;
Subelement is inputted, for the characteristics of image of the 4th sample point described in the current medical image to be input into the second machine
In device learning model, so that second machine learning model exports the 4th sampled point phase described in the current medical image
For the position offset of the first key point;
Subelement is searched, for searching L2 norms most in the position offset from the 4th sampled point relative to the first key point
Small position offset;
4th determination subelement, for by the 4th sampled point corresponding to the position offset of the L2 Norm minimums, as described
Target predicted position of first key point in the current medical image.
9. the device of key point is positioned in a kind of medical image, it is characterised in that described device includes:
Processor, memory and communication bus;
Wherein, the processor is connected with the memory by the communication bus;
The memory is used for store instruction, and the processor, which is used to instruct from the memory calls, to be performed, the instruction bag
Include:
Triggering command in response to positioning key point for current medical image, the selected pixels point from the current medical image,
As the first sampled point;
The characteristics of image of first sample point is input in the first machine learning model, existed to obtain the first key point
The first predicted position in the current medical image, and determine the second sampled point from first predicted position;Wherein, institute
It is based on the characteristics of image of the first history pixel, the first history picture in history medical image to state the first machine learning model
The position offset that vegetarian refreshments is respectively relative to each key point is trained to obtain, and the first history pixel is history medical science
Whole pixels in image;
The characteristics of image of second sample point is input in the second machine learning model, it is crucial to obtain described first
Second predicted position of the point in the current medical image;Wherein, second machine learning model is based on the history
The characteristics of image of second history pixel in medical image, the second history pixel relative to first key point position
Put offset to be trained to obtain, the second history pixel is to set around the first key point described in history medical image
Pixel in neighborhood;
The second predicted position based on first key point in the current medical image, determine that first key point exists
Target predicted position in the current medical image.
10. a kind of computer-readable storage medium, the storage medium is used for store program codes, and described program code is used for right of execution
Profit requires the method that key point is positioned in the medical image described in any one of 1-4.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710931644.6A CN107742312A (en) | 2017-10-09 | 2017-10-09 | A kind of method and apparatus that key point is positioned in medical image |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710931644.6A CN107742312A (en) | 2017-10-09 | 2017-10-09 | A kind of method and apparatus that key point is positioned in medical image |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107742312A true CN107742312A (en) | 2018-02-27 |
Family
ID=61237036
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710931644.6A Pending CN107742312A (en) | 2017-10-09 | 2017-10-09 | A kind of method and apparatus that key point is positioned in medical image |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107742312A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109344710A (en) * | 2018-08-30 | 2019-02-15 | 东软集团股份有限公司 | A kind of localization method of image characteristic point, device, storage medium and processor |
CN111028212A (en) * | 2019-12-02 | 2020-04-17 | 上海联影智能医疗科技有限公司 | Key point detection method and device, computer equipment and storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106340015A (en) * | 2016-08-30 | 2017-01-18 | 沈阳东软医疗***有限公司 | Key point positioning method and device |
CN106530280A (en) * | 2016-10-17 | 2017-03-22 | 沈阳东软医疗***有限公司 | Method and device for locating organ in image |
-
2017
- 2017-10-09 CN CN201710931644.6A patent/CN107742312A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106340015A (en) * | 2016-08-30 | 2017-01-18 | 沈阳东软医疗***有限公司 | Key point positioning method and device |
CN106530280A (en) * | 2016-10-17 | 2017-03-22 | 沈阳东软医疗***有限公司 | Method and device for locating organ in image |
Non-Patent Citations (2)
Title |
---|
YI SUN等: "Deep Convolutional Network Cascade for Facial Point Detection", 《2013 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
陈锐等: "基于级联卷积神经网络的人脸关键点定位", 《四川理工学院学报( 自然科学版)》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109344710A (en) * | 2018-08-30 | 2019-02-15 | 东软集团股份有限公司 | A kind of localization method of image characteristic point, device, storage medium and processor |
CN111028212A (en) * | 2019-12-02 | 2020-04-17 | 上海联影智能医疗科技有限公司 | Key point detection method and device, computer equipment and storage medium |
CN111028212B (en) * | 2019-12-02 | 2024-02-27 | 上海联影智能医疗科技有限公司 | Key point detection method, device, computer equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110276316A (en) | A kind of human body critical point detection method based on deep learning | |
CN107679522A (en) | Action identification method based on multithread LSTM | |
CN107886510A (en) | A kind of prostate MRI dividing methods based on three-dimensional full convolutional neural networks | |
CN106780569A (en) | A kind of human body attitude estimates behavior analysis method | |
CN109902798A (en) | The training method and device of deep neural network | |
CN108447020A (en) | A kind of face super-resolution reconstruction method based on profound convolutional neural networks | |
CN109271933A (en) | The method for carrying out 3 D human body Attitude estimation based on video flowing | |
CN107403141A (en) | Method for detecting human face and device, computer-readable recording medium, equipment | |
CN106067161A (en) | A kind of method that image is carried out super-resolution | |
CN106780543A (en) | A kind of double framework estimating depths and movement technique based on convolutional neural networks | |
CN107644415A (en) | A kind of text image method for evaluating quality and equipment | |
CN107403426A (en) | A kind of target object detection method and equipment | |
CN106920243A (en) | The ceramic material part method for sequence image segmentation of improved full convolutional neural networks | |
CN104504007B (en) | The acquisition methods and system of a kind of image similarity | |
CN107292884A (en) | The method and device of oedema and hemotoncus in a kind of identification MRI image | |
CN107423814A (en) | A kind of method that dynamic network model is established using depth convolutional neural networks | |
US20220335600A1 (en) | Method, device, and storage medium for lesion segmentation and recist diameter prediction via click-driven attention and dual-path connection | |
CN104200240A (en) | Sketch retrieval method based on content adaptive Hash encoding | |
CN110163954A (en) | Three-dimensional house type model generating method, device, equipment and storage medium | |
CN104992403B (en) | Mixed operation operator image redirection method based on visual similarity measurement | |
CN106339753A (en) | Method for effectively enhancing robustness of convolutional neural network | |
CN107507188B (en) | Method and device for extracting image information based on machine learning | |
CN109461177B (en) | Monocular image depth prediction method based on neural network | |
CN109272044A (en) | A kind of image similarity determines method, apparatus, equipment and storage medium | |
CN104036498B (en) | Fast evaluation method of OCT image quality based on layer by layer classification |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
CB02 | Change of applicant information |
Address after: 110179 No. 177-1 Innovation Road, Hunnan District, Shenyang City, Liaoning Province Applicant after: DongSoft Medical System Co., Ltd. Address before: 110179 No. 177-1 Innovation Road, Hunnan District, Shenyang City, Liaoning Province Applicant before: Dongruan Medical Systems Co., Ltd., Shenyang |
|
CB02 | Change of applicant information | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180227 |
|
RJ01 | Rejection of invention patent application after publication |