Classifications

• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V10/00—Arrangements for image or video recognition or understanding
  • G06V10/20—Image preprocessing
  • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
  • G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T7/00—Image analysis
  • G06T7/10—Segmentation; Edge detection
  • G06T7/11—Region-based segmentation
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T7/00—Image analysis
  • G06T7/10—Segmentation; Edge detection
  • G06T7/136—Segmentation; Edge detection involving thresholding
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T7/00—Image analysis
  • G06T7/10—Segmentation; Edge detection
  • G06T7/174—Segmentation; Edge detection involving the use of two or more images
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T7/00—Image analysis
  • G06T7/10—Segmentation; Edge detection
  • G06T7/187—Segmentation; Edge detection involving region growing; involving region merging; involving connected component labelling
• • 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
  • G06T2207/10—Image acquisition modality
  • G06T2207/10072—Tomographic images
  • G06T2207/10081—Computed x-ray tomography [CT]
• • 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
  • G06T2207/10—Image acquisition modality
  • G06T2207/10072—Tomographic images
  • G06T2207/10088—Magnetic resonance imaging [MRI]

Definitions

• the present invention concerns a method for the segmentation of a discrete 3D grid.
• discrete 3D grid an ordered stack of images is intended, where all images have the same resolution and have been captured along the axis perpendicular to the image plane.
• the discrete 3D grid can be obtained, for example, by a scanner device such as, but not limited to, Magnetic Resonance Imaging (MRI) or Computerized Tomography (CT) or other known scanner.
• MRI Magnetic Resonance Imaging
• CT Computerized Tomography
• the discrete 3D grid is stored in a computer memory and the method described herein operates on such 3D grid.
• the method described in this document is not restricted to medical images, but can be used on all data given in a discrete 3D grid format, e.g. 3D-microscopie, geological scans, and other technical fields.
• MRI Magnetic Resonance Imaging
• CT Computerized Tomography
• the scanning devices use different means of scanning a human body and return a stack of 2D-grayscale-images. This stack allows medical professionals to non-invasively observe the scanned patient to identify problems and plan surgeries.
• a DICOM dataset is made up of 2D images, that provide a grayscale image through the scanner plane, with each image being captured at a certain distance along the axis perpendicular to that plane.
• the gray (or Hounsfield) value of an image point is determined by the interaction of the patient's body tissue at the position related to that image point. Since different material densities or consistencies lead to different interactions, the image shows different tissues in different gray values, allowing a medical professional to "see" the inside of the scanned patient.
• Medical image segmentation is used to classify the contents of the image. The goal is, to identify different tissue types and organs. The segmentation can then be used for different purposes, like creating a 3D-model for the scanned region.
• a significant problem in automatic medical image segmentation is, that the Hounsfield values are not identical within the same organ or tissue. This is due to noise created by the scanning device and differences in the tissue itself, which is not homogeneous within organs.
• Another reason for inhomogeneous Hounsfield values that can especially be observed in the bones is, that the bones do not have equal density among their surface. This is especially the case where bones are worn thin (for example at joints) or in patients with a fracture. Especially in the case of fractures, the patient is putting as little stress as possible on the injured bone, which in turn loses density.
• the Hounsfield value alone cannot be used to extract a reliable 3D-model, although it is used in practice to get at least a crude segmentation through thresholding.
• An object of the present invention is to create a method that is able to identify border regions between different tissue types and automatically generate the outlines of surfaces within the scanned data.
• a further aim of this invention is to solve the above mentioned problems in a rational and economical way.
• the above aims are obtained by a method for the segmentation of a discrete 3D grid, the discrete 3D grid being created from an ordered stack of images acquired by a scanning device in a single series, the method comprising at least the following steps:
• N c ⁇ N j ⁇ T 2 is verified, wherein N c is the normal vector of point V c , N j is the normal vector of point V j and T 2 is a predefined threshold until all the neighbor points V j of the current point V c have been considered and then erase the current point V c from the queue Q;
• the algorithm described in the present patent application is able to identify border regions between different tissue types and to automatically generate the outlines of surfaces within the scanned data.
• the present invention does not require any pre-processing of the data, nor does it require manual user interaction. Since the algorithm works directly on the gradients of the measured Hounsfield values, it works simultaneously with both MRI and CT data.
• the result of running the algorithm on data acquired by a medical scanning device is a set of separate 3D objects, each representing an object within the scanned region, e.g. one object for each disjoint bone or even bone fragment in case of broken bones present in the data, organs, muscles and skin.
• a further step of pre- selecting points lying near surfaces of the computer generated image and discarding points belonging to the same tissue can be performed, the preselecting step comprising the calculation of a number G i for each point V i , said number G i being representative of the number of point neighbors V j of the point V i having similar surface normal to the point V i .
• FIG. 1 illustrates a first step of a normal computation in 2D and a neighborhood of 1;
• Figure 2 illustrates the vectors from Figure 1 multiplied by the difference in the Hounsfield value of the point corresponding to the respective vector
• Figure 6a shows a single layer of a Digital Imaging and Communications in Medicine (DICOM).
• Figures 6b-6c show the different values computed for this slice and Figure 6d shows the resulting final segmentation;
• Figure 7a shows an image of the 3D-model generated by the segmentation of two feet of a patient ' s DICOM file
• Figures 7b-7d a 3D-model of the hip bone segmented from another DICOM file
• FIG. 8 shows the gluteal muscles in red on top and the skin of the buttocks of a patient in the lower part of the image.
• the algorithm proposed in the present patent application comprise a series of consecutive steps, depending on the data quality: 1. Compute normal vector in each data point
• DICOM-files are the de facto standard in storing medical image data, acquired by scanners like MRI, CT and similar scanning devices.
• point describes a single pixel element of a single DICOM image.
• grid will be used to describe the 3D-data structure that is created by all the points in a single series within a DICOM-file.
• the grid of a DICOM-file will therefore have the dimensions:
• Depth d number of images in the DICOM (considering a single series only)
• the notation V i is used to refer to an arbitrary point within the grid. Each point V i will initially have a Hounsfield-value of determined by the scanner output. This is the input that the algorithm of the present invention is operating on.
• V [xyz] is used to refer to the point with coordinates x ⁇ [0,w — 1],y ⁇ [0,h- 1],z ⁇ [0,d— 1] inside of the grid.
• the vector is the normalized vector from the center of to the center of V j and is computed using the grid coordinates of V i and
• n neighborhood of V i is then defined as the set of all points with a distance to smaller or equal to n:
• an error value E i and a normal vector can be computed for each grid point V i as follows:
• the following step of pre-selecting points near surfaces can be used.
• points that are not in close proximity to any surfaces can be excluded.
• Step 3 the step of pre-selecting points near surfaces should only be used when the data is very noisy. This step does not have positive effects in smooth data, since for this kind of data Step 3 is enough. Now we will create the surfaces contained within the Grid. To do that, we loop over all V i in the grid that were not discarded in step 2 (in case of noisy data).
• a final step of the method provides for restricting the surfaces identified to lie at the point of strongest contrast.
• the points in a single surface S i are points that share an angle smaller than T 2 with their direct neighbors. But due to the fact that the normal computation in the first step of the present method was carried out over a larger area, the surface is blurred out (see Figure 5 and Figure 6b). Instead of just containing the points where the gradient between different tissue types is strongest, it also contains the points that are close enough to this region.
• the result is an image stack with highly accurate segmentation of all the objects in this stack.
• a 3D-model of the scanned area is required, it is now trivial to construct it by different means.
• the one we used is to simply run a marching cubes algorithm over the whole grid, one surface at a time.For the algorithm to produce a good 3D model, the grid points are set to zero while only marking the points contained in the current surface to be above the threshold set for the marching cubes algorithm. This way, a precise 3D model is created for each surface within the scanned area.
• figure 1 illustrates a first step of a normal computation in 2D and a neighborhood n of 1, figure 1 and the following figures being merely an exemplification of the various possibilities of the method.
• the normal computation in 2D and a neighborhood n of starts with a normalized vector from the current point in the center, to the center of each of its 8 neighbors.
• Figure 2 illustrates the vectors from Figure 1 multiplied by the difference in the Hounsfield value of the point corresponding to the respective vector. Both vectors along the diagonal from top-left to bottom-right vanish, since the Hounsfield-values are equal along that diagonal in this example. Furthermore, three vectors have changed direction, due to the negative sign in the Hounsfield difference.
• Figure 3 shows the result of summing up the vectors from Figure 2.
• Figure 4 shows the resulting normal vector for the center pixel.
• Figure 5 shows the resulting vectors when applying the above described computation to every point in the given grid.
• Figures 6a-6d shows various images related to a slice of Digital Imaging and Communications in Medicine (DICOM).
• DICOM Digital Imaging and Communications in Medicine
• the image of Figure 6a shows the Hounsfield values in a slice of a DICOM.
• the image of Figure 6b shows the normal vectors computed from that data set.
• the 3D-vector is encoded in the RGB-space of the color-image.Thiswas done by taking each component of the vector, adding 1. 0 and multiplying it by 127. Since each component of the normalized vector is in the range [-1.0, 1.0], each color component of the resulting image will therefore be in the range [0, 254].
• the image visualizes the regularity of the vectors along borders used in the growing of the surfaces.
• the image of Figure 6c shows the error-value. Red pixels have an error value between -1.0 and -0.5 while blue pixels have an error value between 0.5 and 1.0. Green pixels show error values between -0.5 and 0.5.
• the image of Figure 6d shows the resulting segmentation using both the normal vectors for surface growth as well as the error values for border identification.
• Figure 7a shows on the left side of the image the segmentation of two feet of a patient segmented from one DICOM image
• Figures 7b- 7c show the hip bone segmented from another DICOM image.
• Figure 8 This image of Figure 8 was extracted from the same DICOM that produced the hip bone in Figures 7a-c.
• the algorithm described herein is capable of automatically segmenting all parts at once, while separating the different surfaces into different 3D-models.
• Figure 8 also demonstrates the extreme precision of the algorithm, showing every dent in the skin of the patient (lower part) and the clearly distinct muscle fibers.
• a preferred way to perform the computation of the described method is to compute the normal vectors in the 3D grid, which gives better results than performing this computation in each of the 2D images of the stack separately.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• General Physics & Mathematics (AREA)
• Theoretical Computer Science (AREA)
• Computer Vision & Pattern Recognition (AREA)
• Multimedia (AREA)
• Apparatus For Radiation Diagnosis (AREA)
• Image Processing (AREA)
• Image Analysis (AREA)

Applications Claiming Priority (2)

Publications (1)

Family

ID=75143680

Family Applications (1)

Country Status (2)

Cited By (1)

Citations (4)

Family Cites Families (2)

• 2020
  • 2020-03-09 CH CH000269/2020A patent/CH717198B1/it unknown
• 2021
  • 2021-03-08 WO PCT/IB2021/051897 patent/WO2021181238A1/en active Application Filing

Patent Citations (4)

Also Published As

Similar Documents

Legal Events