CN115641345A - Multiple myeloma cell morphology fine segmentation method based on deep learning - Google Patents
Multiple myeloma cell morphology fine segmentation method based on deep learning Download PDFInfo
- Publication number
- CN115641345A CN115641345A CN202211271668.0A CN202211271668A CN115641345A CN 115641345 A CN115641345 A CN 115641345A CN 202211271668 A CN202211271668 A CN 202211271668A CN 115641345 A CN115641345 A CN 115641345A
- Authority
- CN
- China
- Prior art keywords
- multiple myeloma
- attention
- segmentation
- training
- network
- 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
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a multiple myeloma cell morphology fine segmentation method based on deep learning, which adopts a network structure of an encoder and a decoder, adopts a series-parallel multi-attention residual error structure in an encoder characteristic extraction part, and introduces multi-scale deep supervision and focus loss in a decoder part, so that the network extracts more abundant channel, local space and global space characteristics, the multi-scale deep supervision can be fused with different scale outputs, and the multiple myeloma cells which are less in quantity and difficult to segment are concerned more. The method overcomes the defect of difficult fine segmentation of the cellular morphology of the multiple myeloma, has the advantages of automation, low cost, high efficiency and the like, and has great significance for early medical diagnosis and prognosis evaluation.
Description
Technical Field
The invention belongs to the technical field of biomedical image processing, and particularly relates to a deep learning-based multiple myeloma cell morphology fine segmentation method.
Background
Bone marrow cell morphology analysis is an important task in the diagnosis of hematological malignancies, and is usually performed in clinical hematological tests, such as detecting leukemia, multiple myeloma, lymphoma, bone marrow invasion, etc. The plasma cell morphology observed in bone marrow smears cannot be separated from the diagnostic and prognostic assessment of multiple myeloma, and can be an independent prognostic factor for multiple myeloma. The cytomorphological specialist classifies and lists the cells by observing smear Switzerland staining of bone marrow and peripheral blood cells under a common optical microscope, checks the presence or absence of abnormal plasma cells, and roughly classifies primary plasma cells, immature plasma cells and mature plasma cells to judge whether malignant plasma cells (myeloma cells) exist. However, even after years of experience by experts has accumulated, considerable effort is required to analyze the fine segmentation of myeloma cells, and visual inspection does not give quantitative data on the morphology of individual myeloma cells. Therefore, it is urgent to develop an effective automatic analysis method.
The morphological identification and enumeration of myeloma cells comprises two main stages: and (4) identifying and segmenting. In these steps, myeloma cell segmentation is very important because it affects the performance of the final segmentation parameters. Particularly, in the actual microscopic image of myeloma cells, the areas of the nucleus and cytoplasm cannot be artificially measured, and the nuclear-to-cytoplasmic ratio cannot be calculated. After the myeloma cells are segmented into the nucleus and the cytoplasm, the areas of the nucleus and the cytoplasm can be easily calculated, and quantitative data of the morphology of each myeloma cell can be further calculated, so that high-efficiency and rapid prognostic analysis can be performed. Therefore, research has focused on breaking through the limitations of segmentation of myeloma cells through the nucleus and cytoplasm, and expanding the scale of cost-effective medical and biomedical research.
The medical image segmentation is a complex and key step in the field of medical image processing and analysis, and aims to segment parts with certain special meanings in a medical image and extract related features, so that the change of anatomical or pathological structures in the image is clearer, a reliable basis is provided for clinical diagnosis and pathological research, and a doctor is assisted to make more accurate diagnosis. Due to the complexity of the medical image, a series of problems such as non-uniformity, individual difference and the like need to be solved in the segmentation process, so that the general image segmentation method is difficult to be directly applied to medical image segmentation. Currently, medical image segmentation is still evolving from manual segmentation or semi-automatic segmentation to fully automatic segmentation.
Early methods of medical image segmentation generally relied on edge detection, template matching techniques, statistical shape models, active contours, and traditional machine learning techniques. These methods achieve good results to some extent, but due to the difficulty of feature representation, image segmentation remains one of the most challenging issues in the field of computer vision. Especially, feature extraction of medical images is more difficult than that of ordinary RGB images because the former often has the problems of blurring, noise, low contrast, and the like. Due to the rapid development of the deep learning technology, the medical image segmentation does not need the characteristics made by hands, and the convolutional neural network successfully realizes the hierarchical characteristic representation of the image, so that the convolutional neural network becomes the most popular research topic in the field of image processing and computer vision.
In recent years, although some progress has been made in image segmentation tasks from early Full Convolution Networks (FCN) and U-Net to current Pyramid Scene analysis Networks (PSPNet) and deep lab, these methods are prone to erroneous recognition and erroneous recognition in distinguishing specific tumor cells from general tumor cells and from general cell segmentation, and are not good in boundary segmentation, and are not balanced between stained cells and background regions.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a multiple myeloma cell morphology fine segmentation method based on deep learning, which comprises the following steps:
step 2, performing data enhancement on the gold standard data set constructed in the step 1;
step 3, constructing a multiple myeloma cell morphology fine segmentation network based on deep learning;
step 4, training the multiple myeloma cell morphology fine segmentation network constructed in the step 3 by using the gold standard data set subjected to data enhancement in the step 2;
and 5, finely dividing the nucleus and cytoplasm of the multiple myeloma cell by using the network model trained in the step 4.
In step 1, the collected normal cell dataset is labeled with segmentation labeling software as a first gold standard dataset, and the collected multiple myeloma cell dataset is labeled with segmentation labeling software as a second gold standard dataset.
In addition, in the step 2, two data enhancement modes of generating sample data of an anti-network and enhancing mosaic are introduced except that common enhancement modes including turning, translation, rotation, scaling, clipping, sharpening and color transformation are adopted. And the generation of the antagonistic network can generate multiple myeloma cell images with more forms and numbers, and the diversity of sample data is enriched. The mosaic enhancement scales and splices a plurality of cell images into one multiple myeloma cell image, so that the number of multiple myeloma cells in each sample can be increased, the imbalance among target cells, non-target cells and a background is reduced, the diversity of the sample can be increased, and the network performance is improved.
Moreover, the network in step 3 adopts an encoder and a decoder structure, the encoder part adopts a series-parallel multi-attention residual structure, and the two series-parallel double-attention structures are connected in parallel, so that the context information of local space, non-local space, channel and space channel can be effectively extracted. The serial structure is a channel attention serial branch and a space attention serial branch respectively, rich channel and local space features can be extracted, the channel attention serial branch and the non-local attention serial branch can extract the rich channel and the non-local space features, then the two serial double-attention structures are connected in parallel, and the feature extraction richness can be further improved. The channel attention module performs maximum pooling and average pooling on the input feature map respectively to form two weight vectors, then performs mapping to form the weight of each channel through the same shared full-connection layer respectively, adds the weights, outputs a normalized channel attention weight feature map through a Sigmoid function, and finally multiplies the normalized channel attention weight feature map by the original feature map according to the channel to finish the recalibration of the original feature by the channel attention. The space attention module performs maximum pooling and average pooling on the input feature maps according to channels, the obtained two feature maps are stacked to form a feature map space weight, the feature map space weight is then upgraded to an original dimension after one-time convolution layer, and finally the feature map subjected to Sigmoid activation function standardization processing is merged with the original input feature map to finish the recalibration of the space attention on the original features. Different from the spatial attention, the local relation is obtained through a local area, the non-local attention is that the response of a certain pixel point is the sum of the feature weights of all other points, each point is associated with all other points, long-distance dependence is captured, the non-local spatial idea is realized, the structure is that the input feature maps are respectively subjected to linear mapping through 1 multiplied by 1 to obtain three feature maps, after dimension transformation operation, point multiplication operation is carried out on two feature maps, then the weight coefficient of the self-attention is obtained through a normalized exponential function, point multiplication is carried out on the weight coefficient and the third feature map obtained through previous convolution, finally, residual operation is carried out on the three feature maps and the original input feature map to obtain the non-local attention feature map, and the recalibration of the non-local attention on the original feature is completed. And a multi-scale deep supervision strategy is introduced into a decoder part to fuse the segmentation results of the shallow and deep features and improve the segmentation and identification capabilities of large targets and small targets.
On the loss function, focus loss is introduced to effectively solve the problem of sample class imbalance, so that the network focuses more on the less number of difficult myeloma cells in the picture. The Loss function adopts a multi-classification Dice Loss and Focal Loss weighted summation mode, and the specific calculation mode is as follows:
in the formula, TP p (c)、FN p (c)、FP p (c) True positive, false negative, false positive for category c, respectively; p is a radical of n (c) A probability of predicting as class c for the nth sample; g n (c) Denotes the n-thTrue value when sample is class c; c represents the total number of whole sample set categories (including background classes); λ is L Dice Loss and L Focal The balance factor of the loss can be set to be 0.1, 0.5 or 1 according to the performance index on the test set; α and β are balance factors that penalize false negatives and false positives; n is the number of the entire sample set.
Moreover, the pre-training and migration learning strategy is adopted in the step 4, and the first gold standard dataset of the common cells is firstly expressed as N 1 :N 2 Dividing the ratio into a training set and a testing set at random, amplifying the training set by data enhancement, pre-training on the multiple myeloma cell morphology fine segmentation network provided by the invention until the segmentation index of the testing set is highest than the IOU, stopping training to obtain a pre-training model, and then carrying out second gold standard data set on tumor cells of the multiple myeloma according to N 1 :N 2 And randomly dividing the ratio into a training set and a testing set, carrying out transfer learning on the pre-training model, and carrying out fine-tuning training on the amplified multiple myeloma cell training set. During fine-tuning training, weights except the prediction head in the network are frozen, and N training is carried out 3 After the iteration, the frozen network weight is removed, and the whole network is trained N 4 And (5) iterating for the second time until the intersection of the segmentation indexes on the gold standard data test set of the multiple myeloma cells is higher than the IOU, and stopping training to obtain the segmentation model suitable for the multiple myeloma cells.
Furthermore, in the step 5, a stained bone marrow cell image is given, and after the stained bone marrow cell image is subjected to horizontal overturning, vertical overturning, rotating angle, rotating and overturning operations, N is obtained 5 Respectively carrying out reasoning and prediction on the transformed cell images with different visual angles by using the trained network model in the step 4, and then carrying out N 5 And fusing and averaging the segmentation results to obtain the final high-confidence fine segmentation image of the multiple myeloma cell nucleus and the cytoplasm.
Compared with the prior art, the invention has the following advantages:
1) The method has the advantages that the artificial setting of features is not needed, the deep learning neural network is introduced to realize the automatic segmentation of the multiple myeloma cell nucleus and cytoplasm, the multiple myeloma cell nucleus and cytoplasm area is extracted, the nuclear-cytoplasmic ratio is further calculated to help morphological fine segmentation, and the method is put forward for the first time in a task of morphological fine segmentation of the multiple myeloma cell.
2) The semantic segmentation neural network adopts an encoder and a decoder structure, a brand-new series-parallel multi-attention residual error structure is designed in an encoder part for feature extraction, channel attention and space attention are connected in series, and channel attention and non-local attention are connected in series, so that context information of a local space, a non-local space, a channel and a space channel can be effectively extracted, and the context information and the non-local space, the channel and the space channel are integrated in parallel, and the richness of feature extraction can be further improved.
3) A multi-scale deep supervision strategy is added in a decoder part, and the segmentation results of shallow and deep features are fused, so that the segmentation and identification capabilities of large targets and small targets can be remarkably improved.
4) In the aspect of loss function design, the focus loss is added, so that the problem of unbalanced sample types can be effectively solved, and the network focuses more on the myeloma cells which are less in number and difficult to segment.
5) And in the model inference prediction stage, data enhancement operations such as horizontal turning, vertical turning, rotation angle and the like are carried out on the original picture to obtain 8 transformed cell images with different viewing angles, inference prediction is carried out respectively, and then 8 segmentation results are fused and averaged to obtain a final output result, so that the segmentation precision and generalization capability are improved, and the form fine segmentation is assisted more accurately and efficiently.
Drawings
FIG. 1 is a flow chart of an embodiment of the present invention.
Fig. 2 is a data set calibration process of the present invention, in which fig. 2 (a) shows multiple myeloma cells artificially labeled by segmentation labeling software, and fig. 2 (b) shows labeled multiple myeloma cell mask images automatically generated by segmentation labeling software after the segmentation labeling is performed manually.
FIG. 3 is a diagram of the structure of each sub-attention in series-parallel multi-attention of the present invention.
FIG. 4 is a network model inference process, according to an embodiment of the invention.
Fig. 5 shows experimental results of the present invention, in which fig. 5 (a) is an original image of multiple myeloma, fig. 5 (b) is a mask image of multiple myeloma cells labeled by segmentation labeling software, and fig. 5 (c) is a mask image of multiple myeloma cells predicted by a network.
Detailed Description
The invention provides a method for finely segmenting multiple myeloma cell morphology based on deep learning, and the technical scheme of the invention is further explained by combining the accompanying drawings and an embodiment.
As shown in fig. 1, the process of the embodiment of the present invention includes the following steps:
Because the data of multiple myeloma cells is few and difficult to collect, a transfer learning method is adopted, and a large number of common cell pictures are collected to be used as auxiliary data to perform model pre-training. The collected general cell data comprises bladder cells, liver cells, blood cells and the like, and is labeled by professional segmentation labeling software labelme to serve as a first gold standard data set. The collected multiple myeloma cell dataset was labeled with labelme as the second gold standard dataset. Fig. 2 shows the process of constructing the gold standard dataset of the present invention, and fig. 2 (a) shows the software label, which outlines the nucleus and cytoplasm of the multiple myeloma cell, and sets the categories, which are indicated by the numbers 0 and 1, respectively. And obtaining the marking information of the stained multiple myeloma image, including contour coordinates and categories, after the software marking is finished. The gold standard data set shown in fig. 2 (b) can be obtained by extracting the contour of the multiple myeloma cell by an image processing technique and performing mask filling on the contour.
And 2, performing data enhancement on the gold standard data set constructed in the step 1.
Besides common enhancement modes including turning, translation, rotation, scaling, clipping, sharpening, color transformation and the like, two data enhancement modes of generating an anti-network to generate sample data and mosaic enhancement are introduced, the anti-network can be generated to generate multiple myeloma cell images with more forms and quantities, the diversity of the sample data is enriched, the mosaic enhancement combines multiple cell images into one multiple myeloma cell image after scaling, the number of multiple myeloma cells in each sample can be increased, the imbalance among target cells, non-target cells and a background is reduced, the diversity of the samples can be increased, and the network performance is improved.
And 3, constructing a multiple myeloma cell morphology fine segmentation network based on deep learning.
By adopting the structures of the encoder and the decoder, a brand-new series-parallel multi-attention residual error structure is designed in the encoder part for feature extraction, and the two series-parallel double-attention structures are connected in parallel, so that the context information of local space, non-local space, channel and space channel can be effectively extracted. The series structure is a channel attention and space attention series branch respectively, and rich channel and local space characteristics can be extracted; the channel attention and the non-local attention are connected in series and branched, rich channel and non-local spatial features can be extracted, and then the two serial double-attention structures are connected in parallel, so that the feature extraction richness can be further improved. The channel attention module performs maximum pooling and average pooling on the input feature map respectively to form two weight vectors, then performs mapping to form the weight of each channel through the same shared full-connection layer respectively, adds the weights, outputs a normalized channel attention weight feature map through a Sigmoid function, and finally multiplies the normalized channel attention weight feature map by the original feature map according to the channel to finish the recalibration of the original feature by the channel attention. The space attention module performs maximum pooling and average pooling on the input feature maps according to channels, the obtained two feature maps are stacked to form a feature map space weight, the feature map space weight is then upgraded to an original dimension after one-time convolution layer, and finally the feature map subjected to Sigmoid activation function standardization processing is merged with the original input feature map to finish the recalibration of the space attention on the original features. Different from the spatial attention, the local relation is obtained through a local area, the non-local attention is that the response of a certain pixel point is the sum of the feature weights of all other points, each point is associated with all other points, long-distance dependence is captured, the non-local spatial idea is realized, the structure is that the input feature maps are respectively subjected to linear mapping through 1 multiplied by 1 to obtain three feature maps, after dimension transformation operation, point multiplication operation is carried out on two feature maps, then the weight coefficient of the self-attention is obtained through a normalized exponential function, point multiplication is carried out on the weight coefficient and the third feature map obtained through previous convolution, finally, residual operation is carried out on the three feature maps and the original input feature map to obtain the non-local attention feature map, and the recalibration of the non-local attention on the original feature is completed. And a multi-scale deep supervision strategy is introduced into a decoder part to fuse the segmentation results of the shallow and deep features and improve the segmentation and identification capabilities of large targets and small targets.
On the Loss function, a Focus Loss (Focus Loss) is introduced to effectively solve the problem of sample class imbalance, so that the network focuses more on a small number of difficult-to-segment myeloma cells in the picture. The Loss function adopts a multi-classification Dice Loss and Focal Loss weighted summation mode, and the specific calculation mode is as follows:
in the formula, TP p (c)、FN p (c)、FP p (c) True positive, false negative, false positive for category c, respectively; p is a radical of n (c) A probability of predicting as class c for the nth sample; g is a radical of formula n (c) A true value when the nth sample is of the class c is shown; c represents the total number of whole sample set categories (including background classes); λ is L Dice Loss and L Focal The balance factor of the loss is determined by the loss,the performance index on the test set can be set to 0.1, 0.5 or 1, and the embodiment is set to 0.5; α and β are balance factors penalizing false negatives and false positives, both set to 0.5 in this example; n is the number of the entire sample set.
And 4, training the multiple myeloma cell morphology fine segmentation network constructed in the step 3 by using the gold standard data set subjected to data enhancement in the step 2.
A pre-training and migration learning strategy is adopted, firstly, a common cell first gold standard data set is randomly divided into a training set and a testing set according to 8:2 proportion, after the training set is amplified through data enhancement, pre-training is carried out on the multiple myeloma cell morphology fine segmentation network provided by the invention until the segmentation index cross-over ratio (IOU) of the testing set is the highest, training is stopped to obtain a pre-training model, then a tumor cell second gold standard data set of multiple myeloma is randomly divided into the training set and the testing set according to 8:2 proportion, the pre-training model is subjected to migration learning, and fine-tuning training is carried out on the amplified multiple myeloma cell training set. And during fine tuning training, freezing the weights except the prediction head in the network, training for 10 times of iteration, then removing the frozen network weights, training the whole network for 290 times of iteration until the cross-over ratio (IOU) of the segmentation indexes on the gold standard data test set of the multiple myeloma cells is the highest, and stopping training to obtain the segmentation model suitable for the multiple myeloma cells.
And 5, finely dividing the multiple myeloma cell nucleus and the multiple myeloma cell cytoplasm by using the network model trained in the step 4.
As shown in fig. 4, a stained bone marrow cell image is given, and is subjected to data enhancement operations such as horizontal inversion, vertical inversion, rotation angles (90 degrees, 180 degrees, 270 degrees), rotation and inversion, so as to obtain 8 transformed cell images with different viewing angles, the segmentation networks trained in step 4 are used for reasoning and prediction, and the 8 segmentation results are fused and averaged, so as to obtain the final high-confidence multiple myeloma cell nucleus and cytoplasm fine segmentation image.
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.
Claims (7)
1. A method for finely segmenting the morphology of multiple myeloma cells based on deep learning is characterized by comprising the following steps:
step 1, collecting sample data and constructing a gold standard data set;
step 2, performing data enhancement on the gold standard data set constructed in the step 1;
step 3, constructing a multiple myeloma cell morphology fine segmentation network based on deep learning; an encoder and a decoder are adopted, the encoder part adopts a series-parallel multi-attention residual error structure, two series double-attention structures are connected in parallel, the series structures are respectively a channel attention and space attention series branch, a channel attention and non-local attention series branch, and a multi-scale deep supervision strategy is introduced into the decoder part to fuse shallow and deep feature segmentation results and improve the segmentation and identification capabilities of a large target and a small target;
step 4, training the multiple myeloma cell morphology fine segmentation network constructed in the step 3 by using the gold standard data set subjected to data enhancement in the step 2;
and 5, finely dividing the nucleus and cytoplasm of the multiple myeloma cell by using the network model trained in the step 4.
2. The method for multiple myeloma cell morphology fine segmentation based on deep learning according to claim 1, wherein the method comprises the following steps: in step 1, the collected common cell data set is labeled by segmentation labeling software to be used as a first gold standard data set, and the collected multiple myeloma cell data set is labeled by segmentation labeling software to be used as a second gold standard data set.
3. The method of claim 1, wherein the method comprises the following steps: step 2, introducing two data enhancement modes of generating sample data of an antagonistic network and enhancing mosaic except for adopting common enhancement modes comprising turning, translation, rotation, scaling, clipping, sharpening and color transformation; the generation of the countermeasure network can generate multiple myeloma cell images with more forms and quantities, and the diversity of sample data is enriched; mosaic enhancement scales and combines multiple cell images into one multiple myeloma cell image, so that the number of multiple myeloma cells in each sample can be increased, imbalance among target cells, non-target cells and a background can be reduced, sample diversity can be increased, and network performance can be improved.
4. The method of claim 1, wherein the method comprises the following steps: in the step 3, the channel attention and the space attention are connected in series, so that rich channel and local space characteristics can be extracted; the channel attention and the non-local attention are connected in series, so that rich channel and non-local spatial features can be extracted; the two serial double-attention structures are connected in parallel, so that the richness of extracted features can be further improved; the channel attention module performs maximum pooling and average pooling on the input feature map respectively to form two weight vectors, then performs mapping to form the weight of each channel through the same shared full-connection layer respectively, adds the weights, outputs a normalized channel attention weight feature map through a Sigmoid function, and finally multiplies the normalized channel attention weight feature map by the original feature map according to the channel to finish the recalibration of the original feature by the channel attention; the space attention module performs maximum pooling and average pooling on the input feature maps according to channels, the obtained two feature maps are stacked to form a feature map space weight, the feature map space weight is then upgraded to an original dimension after being subjected to primary convolution, and finally the feature map subjected to Sigmoid activation function standardization processing is merged with the original input feature map to finish the recalibration of the space attention on the original features; different from the spatial attention, the local relation is obtained through a local area, the non-local attention is that the response of a certain pixel point is the sum of the feature weights of all other points, each point is associated with all other points, long-distance dependence is captured, the non-local spatial idea is realized, the structure is that the input feature maps are respectively subjected to linear mapping through convolution to obtain three feature maps, after dimension transformation operation, point multiplication operation is carried out on two feature maps, then the weight coefficient of the attention is obtained through a normalized exponential function, point multiplication is carried out on the weight coefficient and the third feature map obtained through the previous convolution, finally, residual operation is carried out on the input feature map to obtain a non-local attention feature map, and the recalibration of the non-local attention on the original features is completed.
5. The method of claim 1, wherein the method comprises the following steps: in the step 3, the Loss function adopts a multi-classification Dice Loss and focallloss weighted summation mode, and the specific calculation mode is as follows:
in the formula, TP p (c)、FN p (c)、FP p (c) True positive, false negative, false positive for category c, respectively; p is a radical of n (c) A probability of predicting as class c for the nth sample; g n (c) A true value when the nth sample is of the class c is shown; c represents an integerTotal number of sample set classes, including background class; λ is L Dice Loss and L Focal A balance factor for loss is set to 0.1, 0.5 or 1 according to the performance index on the test set; α and β are balance factors that penalize false negatives and false positives; n is the number of the entire sample set.
6. The method of claim 1, wherein the method comprises the following steps: in step 4, a pre-training and migration learning strategy is adopted, and the first gold standard data set of the common cells is firstly expressed according to N 1 :N 2 Dividing the ratio into a training set and a testing set at random, amplifying the training set by data enhancement, pre-training on the multiple myeloma cell morphology fine segmentation network provided by the invention until the segmentation index of the testing set is highest than the IOU, stopping training to obtain a pre-training model, and then carrying out second gold standard data set on tumor cells of the multiple myeloma according to N 1 :N 2 Randomly dividing the ratio into a training set and a testing set, carrying out transfer learning on the pre-training model, and carrying out fine tuning training on the amplified multiple myeloma cell training set; during fine-tuning training, weights except the prediction head in the network are frozen, and N training is carried out 3 After the iteration, the frozen network weight is removed, and the whole network is trained N 4 And (5) iterating for the second time until the intersection of the segmentation indexes on the gold standard data test set of the multiple myeloma cells is higher than the IOU, and stopping training to obtain the segmentation model suitable for the multiple myeloma cells.
7. The method of claim 1, wherein the method comprises the following steps: giving a stained bone marrow cell image in the step 5, and performing horizontal turning, vertical turning, rotation angle, rotation and turning on the stained bone marrow cell image to obtain N 5 Respectively carrying out reasoning prediction on the transformed cell images with different visual angles by using the trained network model in the step 4, and then carrying out N 5 Fusing and averaging the segmentation results to obtain the final multiple myeloma cell with high confidenceThe nucleus and cytoplasm finely divide the image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211271668.0A CN115641345A (en) | 2022-10-18 | 2022-10-18 | Multiple myeloma cell morphology fine segmentation method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211271668.0A CN115641345A (en) | 2022-10-18 | 2022-10-18 | Multiple myeloma cell morphology fine segmentation method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115641345A true CN115641345A (en) | 2023-01-24 |
Family
ID=84945355
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211271668.0A Pending CN115641345A (en) | 2022-10-18 | 2022-10-18 | Multiple myeloma cell morphology fine segmentation method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115641345A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115984296A (en) * | 2023-03-21 | 2023-04-18 | 译企科技(成都)有限公司 | Medical image segmentation method and system applying multi-attention mechanism |
| CN116309545A (en) * | 2023-05-10 | 2023-06-23 | 湖北工业大学 | Single-stage cell nucleus instance segmentation method for medical microscopic image |
-
2022
- 2022-10-18 CN CN202211271668.0A patent/CN115641345A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115984296A (en) * | 2023-03-21 | 2023-04-18 | 译企科技(成都)有限公司 | Medical image segmentation method and system applying multi-attention mechanism |
| CN115984296B (en) * | 2023-03-21 | 2023-06-13 | 译企科技(成都)有限公司 | Medical image segmentation method and system applying multi-attention mechanism |
| CN116309545A (en) * | 2023-05-10 | 2023-06-23 | 湖北工业大学 | Single-stage cell nucleus instance segmentation method for medical microscopic image |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109523520B (en) | Chromosome automatic counting method based on deep learning | |
| CN115641345A (en) | Multiple myeloma cell morphology fine segmentation method based on deep learning | |
| CN112435243A (en) | Automatic analysis system and method for full-slice digital pathological image | |
| CN112288706A (en) | Automatic chromosome karyotype analysis and abnormality detection method | |
| CN109102498B (en) | Method for segmenting cluster type cell nucleus in cervical smear image | |
| CN113393443B (en) | HE pathological image cell nucleus segmentation method and system | |
| CN112215801A (en) | Pathological image classification method and system based on deep learning and machine learning | |
| CN110796661B (en) | Fungal microscopic image segmentation detection method and system based on convolutional neural network | |
| CN112215217B (en) | Digital image recognition method and device for simulating doctor to read film | |
| CN112990214A (en) | Medical image feature recognition prediction model | |
| CN113393454A (en) | Method and device for segmenting pathological target examples in biopsy tissues | |
| CN115546605A (en) | Training method and device based on image labeling and segmentation model | |
| CN115359264A (en) | Intensive distribution adhesion cell deep learning identification method | |
| CN113902669A (en) | Method and system for reading urine exfoliative cell fluid-based smear | |
| CN116468935A (en) | Multi-core convolutional network-based stepwise classification and identification method for traffic signs | |
| CN110414317B (en) | Full-automatic leukocyte classification counting method based on capsule network | |
| CN115206495A (en) | Renal cancer pathological image analysis method and system based on CoAtNet deep learning and intelligent microscopic device | |
| CN114387596A (en) | Automatic interpretation system for cytopathology smear | |
| Abrol et al. | An automated segmentation of leukocytes using modified watershed algorithm on peripheral blood smear images | |
| CN115131628A (en) | Mammary gland image classification method and equipment based on typing auxiliary information | |
| CN114898866A (en) | Thyroid cell auxiliary diagnosis method, equipment and storage medium | |
| CN114821046B (en) | Method and system for cell detection and cell nucleus segmentation based on cell image | |
| CN113222928A (en) | Artificial intelligent urothelial cancer recognition system for urocytology | |
| CN118015614A (en) | High-risk del (17 p 13)/p 53 gene deletion positive multiple myeloma cell classification method | |
| WO2014053520A1 (en) | Targeting cell nuclei for the automation of raman spectroscopy in cytology |
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 |