Lymphoma histological grading method and system, computer equipment and storage medium
The cell separation and classification of lymphoma pathological section images were solved through the Hover-NET nuclear segmentation model and classifier, and the problem of insufficient consistency and objectivity of manual judgments in the histological grading of follicular lymphoma was solved, achieving efficient and accurate Diagnostic results.
Patent Information
- Application Number
- CN202411236126.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-04
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-09-04
AI Technical Summary
In the prior art, the histological grading of follicular lymphoma relies on manual judgment, and there is insufficient consistency and objectivity, resulting in low diagnostic efficiency and high risk of misdiagnosis and misdiagnosis.
Using the Hover-NET nuclear segmentation model and classifier, cells are separated and classified by lymphoma pathological section images, histological grading results and number of cell types are output, subjective judgment is reduced, and diagnostic accuracy and efficiency are improved.
It significantly improves the diagnostic accuracy and efficiency of pathological grading of follicular lymphoma, reduces the risk of misdiagnosis and misdiagnosis, and enhances the standardization and consistency of grading.
Smart Images

Figure CN119068479A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of pathological grading, and in particular relates to a lymphoma histological grading method, system, computer equipment and storage medium. Background Art
[0002] Follicular lymphoma (FL) is a type of non-Hodgkin's B-cell lymphoma. The incidence of lymphoma has been increasing in recent years and is now among the top ten malignant tumors. The diagnosis of FL mainly relies on histopathological examination. Pathologists determine the histological grade of FL by counting the number of centroblasts in at least 10 high-power fields (HPF). According to the number of centroblasts in each HPF, FL is divided into three histological grades. Grade 1: ≤5 centroblasts per field of view; Grade 2: 6-15 centroblasts; Grade 3: >15 centroblasts. In grade 3, if a small number of centroblasts are observed, it is divided into 3a; if the centroblasts are completely absent and there are a large number of centroblasts, it is grade 3b, of which grades 1 and 2 are considered low-grade and grade 3 is considered high-grade. This histological grading plays a key role in guiding patient treatment plans and prognosis.
[0003] According to existing literature reports, the transformation rate of high-grade FL is significantly increased compared with low-grade FL, which indicates that a higher level of pathological grading may be a predictive indicator of FL histological transformation. Specifically, as the FL grade increases, its invasiveness increases, making it more susceptible to transformation. In particular, grade 3b FL is highly similar to diffuse large B-cell lymphoma in terms of gene expression profile, which further supports the correlation between high-grade FL and histological transformation (Histologic transformation) HT. Therefore, accurate pathological grading is crucial to guide treatment decisions during clinical diagnosis and treatment.
[0004] Its pathological grading mainly depends on the number of centroblasts observed under high-power fields. Pathologists manually count the number of centroblasts under 10 high-power fields (2 mm 2 ) to determine the pathological grade. However, the current histological grading of FL relies on the counting and interpretation of pathologists under the microscope, which is highly subjective, has poor consistency, and has low interpretation efficiency.
[0005] In traditional medical practice, the histopathological grading of FL is mainly based on detailed observation and analysis of tissue sections by pathologists using microscopes. However, the effectiveness of this method is limited by the magnification ability and field of view of the microscope, and usually requires doctors to invest a lot of time to carefully analyze and understand the subtle changes on the sections. The traditional tissue section analysis method relies heavily on the doctor's professional knowledge and clinical experience, which brings a certain degree of subjective judgment. Therefore, even for experienced pathologists, there may be differences in the diagnostic results of the same tissue section, which highlights the potential limitations of this method in terms of consistency and objectivity.
[0006] With the rapid development of advanced imaging systems and WSI scanner technology, the field of pathology is undergoing a major transformation. Traditional microscope reading methods are gradually being replaced by digital diagnosis based on WSI. In this round of technological revolution, computer-assisted analysis has shifted from relying mainly on qualitative judgment to more accurate quantitative analysis. The efficient integration and analysis of large amounts of data using DL models not only improves the efficiency of pathological diagnosis, but also makes the diagnostic process more objective and standardized, significantly improving the consistency of results. This method effectively reduces the differences in subjective judgments between different pathologists during the diagnostic process. At present, artificial intelligence has been applied to differential diagnosis, subtype identification, grading, and prognosis assessment. Using artificial intelligence methods for digital pathology processing can be comparable to experienced pathologists in terms of diagnostic accuracy and can meet users' higher performance requirements. These advances not only represent major advances in medical technology, but also provide new possibilities for future pathological diagnosis.
[0007] In the field of pathology research, artificial intelligence technology, especially deep learning algorithms, has become an important tool to provide diagnostic accuracy and efficiency. At present, DL models have been successfully applied to many aspects such as automated diagnosis of tumors, quantitative evaluation of images, and detailed analysis of features, which indicates the potential of artificial intelligence technology in improving the quality of tumor diagnosis and treatment and the depth of research. However, due to the diversity of lymphoma classification and the complexity of diagnosis, the application of artificial intelligence in the field of lymphoma pathology diagnosis is still relatively small. Summary of the invention
[0008] The object of the present invention is to provide a lymphoma histological grading method, system, computer device and storage medium, which solves the problem of potential limitations in terms of consistency and objectivity of relying on manual judgment.
[0009] The present invention is achieved through the following technical solutions:
[0010] The present invention discloses a lymphoma histological grading method, comprising the following steps:
[0011] S1. Screen multiple regions of interest rich in diagnostic information in lymphoma pathological slide images;
[0012] S2, input the region of interest into the trained Hover-NET nuclear segmentation model and output the cell segmentation image;
[0013] Among them, the cell segmentation image can mark the areas of different cell types, the location of each cell and its classification information;
[0014] S3. Input the cell segmentation image into the classifier and output the histological grading results and the number of each cell type.
[0015] Furthermore, in S1, the tool used to screen multiple regions of interest rich in diagnostic information is the OpenHI annotation system.
[0016] Furthermore, in S2, the training process of the Hover-NET core segmentation model is:
[0017] 2.1. Obtain pathological section images of lymphoma with typical morphology and clinical diagnosis of grade 3b as the original image set of centroblasts;
[0018] Acquire morphologically typical pathological section images of clinically diagnosed low-grade lymphoma, and some lightly stained low-grade lymphoma pathological section images, to form a central cell original image set;
[0019] The original image set of centroblasts and the original image set of centroblasts constitute the original data set;
[0020] 2.2. Label the original data set, mark the effective tumor area and the interference cell area, and obtain the labeled data set;
[0021] 2.3. Crop the effective tumor area of each image in the labeled data set to obtain multiple image blocks;
[0022] The image blocks obtained from the pathological section images of 2.4 and 3b grade lymphomas constitute the central mother cell dataset, and the image blocks obtained from the pathological section images of low-grade lymphomas constitute the central cell dataset. The two datasets are divided into training set, validation set and test set. The Hover-NET nuclear segmentation model is trained using the training set until the validation set is qualified and the test set meets the requirements. In this case, the Hover-NET nuclear segmentation model is qualified.
[0023] Furthermore, in 2.2, the tool used to annotate the original data set is the OpenHI annotation system.
[0024] Furthermore, in S3, the classifier can identify three types of cell image patches, which are divided into centroblasts, centroblasts, and other cells;
[0025] The preset threshold in the classifier is the benchmark value of the number of centroblasts;
[0026] After the classifier completes the statistics of the number of central lymphocytes, it compares it with the preset threshold; when the number of central lymphocytes exceeds the threshold, the image is classified as high-grade lymphoma tissue; otherwise, the image is classified as low-grade lymphoma tissue.
[0027] Further, when the image is classified as high-grade lymphoma tissue, the output number is 1;
[0028] When the image is classified as low-grade lymphoma tissue, the output is the number 0.
[0029] The present invention also discloses a lymphoma histological grading system, comprising:
[0030] A screening module, used for screening multiple regions of interest rich in diagnostic information in lymphoma pathological slice images;
[0031] The nuclear segmentation module is used to input the region of interest into the trained Hover-NET nuclear segmentation model and output the cell segmentation image;
[0032] Among them, the cell segmentation image can mark the areas of different cell types, the location of each cell and its classification information;
[0033] The output module is used to input the cell compartment image into the classifier and output the histological grade results and the number of each cell type.
[0034] The present invention also discloses a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the lymphoma histological grading method when executing the computer program.
[0035] The present invention also discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the lymphoma histological grading method are implemented.
[0036] The present invention also discloses a computer program product, comprising a computer program / instruction, which implements the steps of the lymphoma histological grading method when executed by a processor.
[0037] Compared with the prior art, the present invention has the following beneficial technical effects:
[0038] The present invention discloses a method for histological grading of lymphoma, firstly screens out a region of interest, inputs the region of interest into a Hover-NET nuclear segmentation model, outputs a cell segmentation image, and can mark the region of different cell types, the position of each cell and its classification information in the cell segmentation image; however, there is no direct output result, so a classifier is also designed, and the cell segmentation image is input into the classifier, and the histological grading result and the number of each cell type can be output through the classifier. The present invention is to improve the diagnostic accuracy and efficiency of the pathological grading of follicular lymphoma, and the pathological section images of lymphomas with different pathological gradings are efficiently identified by the Hover-NET nuclear segmentation model. The development of this model can provide an accurate classification basis for pathologists, aiming to reduce their workload, and significantly reduce the risk of misdiagnosis and missed diagnosis caused by personal subjective judgment, thereby improving the standardization of follicular lymphoma grading.
[0039] Different from previous Patch studies, this study focused on identification at the cellular level and used the Hover-NET nuclear segmentation model;
[0040] Through its application in lymphoma pathological histological grading, a new path has been explored for the application of image-based machine learning models in the field of medical image analysis.
[0041] Furthermore, when constructing the dataset, the dataset of centroblasts uses pathological slice images of low-grade lymphoma, and the original image set of centroblasts uses 3b high-grade follicular lymphoma. Because the cells in 3a-grade FL are generally more mixed, they were not selected.
[0042] When checking the valid tumor area, the area of interfering cells will also be marked to facilitate the subsequent Hover-NET nuclear segmentation model training.
[0043] Furthermore, the classifier can identify three types of cell numbers, but only focuses on the number of centroblasts when making grading judgments. When the number of centroblasts exceeds the threshold, the image is classified as high-grade lymphoma tissue; otherwise, the image is classified as low-grade lymphoma tissue. Outputting three types of cell numbers is to allow doctors to understand the specific situation of lymphoma tissue more clearly. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 is a flowchart of a lymphoma histological grading method of the present invention;
[0045] Figure 2 The training process of the Hover-NET nuclear segmentation model of the present invention;
[0046] Figure 3 This is the ROI area annotation result for grade 3b follicular lymphoma;
[0047] Figure 4 The results of annotation of tumor follicles; Figure a is a schematic diagram of a digital slice of FL; Figure b is a schematic diagram obtained after the annotator used annotation lines to mark out all tumor follicles;
[0048] Figure 5 is the cropping result; Figure a is an enlarged picture of one of the tumor follicles; Figure b is the cropping process of the enlarged tumor follicle, which is cropped into image blocks of 512x512 pixels for training;
[0049] Figure 6 The figure is a schematic diagram of the image changes in the low-level FL identification process, where green represents the central cell; Figure A is the input low-level lymphoma pathological section image; Figure B is a typical area, and Figure C is an enlarged image of a typical area in Figure B; Figure D is the output cell segmentation image;
[0050] Figure 7 The figure is a schematic diagram of the image changes in the high-level (3a) FL identification process. Red represents centroblasts, blue represents non-tumor cells, and green represents centroblasts. Figure A is the input high-level lymphoma pathology section image; Figure B is a typical area, and Figure C is an enlarged image of a typical area in Figure B; Figure D is the output cell segmentation image;
[0051] Figure 8 The figure is a schematic diagram of the image changes in the high-level (3b) FL identification process, where red represents centroblasts and blue represents non-tumor cells; Figure A is the input high-level lymphoma pathology section image; Figure B is a typical area, and Figure C is an enlarged image of a typical area in Figure B; Figure D is the output cell segmentation image;
[0052] Fig. 9 The performance curves of binary classification of graphic blocks in the test of three datasets are shown;
[0053] Fig.10 Output images of the classifier, where a is a cell segmentation image of high-level FL; b is a cell segmentation image of low-level FL. DETAILED DESCRIPTION
[0054] In order to make the purpose, technical solution and advantages of the present invention more clear, the following is further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the embodiments described are only part of the embodiments of the present invention, not all embodiments.
[0055] The components described and shown in the drawings and embodiments of the present invention may be arranged and designed in various configurations. Therefore, the detailed description of the embodiments of the present invention provided in the following drawings is not intended to limit the scope of the claimed invention, but merely represents a selected embodiment of the present invention. Based on the drawings and embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.
[0056] It should be noted that the terms "comprises", "includes" or any other variations are intended to cover non-exclusive inclusion, so that a process, element, method, article or equipment that includes a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to the process, element, method, article or equipment.
[0057] like Figure 1 As shown, the present invention discloses a lymphoma histological grading method, comprising the following steps:
[0058] S1. Screen multiple regions of interest rich in diagnostic information in lymphoma pathological slide images;
[0059] S2, input the region of interest into the trained Hover-NET nuclear segmentation model and output the cell segmentation image;
[0060] Among them, the cell segmentation image can mark the areas of different cell types, the location of each cell and its classification information;
[0061] S3. Input the cell segmentation image into the trained classifier and output the histological grading results and the number of each cell type.
[0062] When training the Hover-NET core segmentation model, the present invention includes the following parts.
[0063] 1. Selection of clinical sections
[0064] The criteria are as follows: (1) a clear diagnosis of follicular lymphoma or reactive lymph node hyperplasia, with no other concurrent diagnoses; (2) complete tissue sections, excluding puncture biopsy samples with insufficient tissue volume; (3) good section quality; (4) relatively complete clinical and laboratory test data information in the electronic medical record. The microscopic diagnosis of all selected cases was confirmed by two senior pathologists, and difficult-to-diagnose cases were evaluated by a third professional lymphoma pathologist. In addition, the pathological histological grading standard of FL was based on the "Guidelines for the Diagnosis and Treatment of Follicular Lymphoma in China".
[0065] 2. Acquisition of full digital slice image (WSI)
[0066] In this study, the Jiangfeng KF-PRO-005 scanner was used to scan the H&E-stained histopathological sections in all experimental samples. During the scanning process, the objective magnification was adjusted to 40 times to ensure comprehensive coverage of all key staining areas on the sections. The scanning system uses its high-precision algorithm to automatically identify and locate multiple clear focus points on the sections to ensure image clarity and accuracy. When encountering unsatisfactory focus points, the operator will intervene and manually adjust to supplement and improve the focus point settings. The scanner starts to automatically focus and scan step by step, continuously capturing a series of partially overlapping 40-fold magnified square image areas. These individual image blocks are then accurately and synchronously stitched by the scanner's built-in software to construct a complete and detailed image of the slice morphological information, which is called a focus map. After the scanning and stitching process is completed, the resulting focus map is stored in SVS format to form a complete whole digital slide image (WSI). Each WSI is strictly named and classified according to its corresponding slice number and diagnosis name to facilitate subsequent data retrieval and analysis.
[0067] 3. Construct three different datasets
[0068] Dataset 1: Pathological section images of lymphoma with typical morphology and clinical diagnosis of grade 3b were obtained as the original image set of centroblasts;
[0069] The films used for training sets were of high quality and morphologically typical. Since the cells in FL of level 3a are generally more mixed, they were not selected when selecting the training set.
[0070] Acquire pathological section images of clinically diagnosed low-grade lymphoma with typical morphology as the original image set of central cells;
[0071] The original image set of centroblasts and the original image set of centrocytes constitute Dataset 1.
[0072] Dataset 2: Figure 2 As shown in FIG. 1 , based on Dataset 1, some lightly stained low-grade lymphoma pathological section images are added to the original central cell image set as an expanded data set.
[0073] Dataset 3: Based on Dataset 2, the training set data is color normalized.
[0074] The cell nucleus segmentation model was used to test the above three data sets. The recognition accuracy of high / low-grade lymphoma pathological section images is shown in Tables 1 to 3.
[0075] Table 1
[0076]
[0077] The WHO classification standard (150) is a binary discrimination performance test variable, see Fig. 9 , which shows the accuracy of this data set 2, which is a statistical method. The area under the receiver operating characteristic curve (ROC) (AUC) provides a more accurate performance evaluation standard. If the AUC value is between 0.5 and 0.7, the predictive performance of the cell nucleus segmentation model is considered poor; an AUC value of 0.7 to 0.8 indicates that the model has a moderate level of predictive ability; 0.8 to 0.9 indicates that the model has good predictive performance; and an AUC value greater than or equal to 0.9 indicates that the predictive performance of the model has reached an excellent level. This figure is to illustrate that the predictive ability of the cell nucleus segmentation model of the present invention is very good, as shown in Table 3.
[0078] Table 2
[0079]
[0080] Table 3
[0081]
[0082] Through the above data analysis, it is found that the Hover-NET nuclear segmentation model cannot recognize the pathological section images of 1-2 levels of FL with relatively light staining. Therefore, the present invention adds a part of 1-2 levels of FL pathological section images with light staining to the data set. After learning this part of the slices, the Hover-NET nuclear segmentation model has an improved recognition effect on low-level FL.
[0083] When data set 2 was modeled and analyzed using the number of centroblasts as the test variable, the optimal cutoff value was 159, which was very close to the cutoff value of 150 recommended by the WHO guidelines, further confirming the predictive accuracy of data set 2.
[0084] Therefore, Dataset 2 was used as the dataset for training the cell nucleus segmentation model. In the research design process, a different approach was taken from most other studies: the slices were not uniformly restained. This decision was based on the consideration of the generalization ability of the model. It was believed that by using original slice samples from the real world instead of artificially processed samples, the dataset of the present invention could be more adaptable and generalizable.
[0085] 4. Image Annotation
[0086] The data set is labeled to mark the effective tumor area and the interference cell area to obtain a labeled data set.
[0087] like Figure 3As shown, after annotating one of the pathological slice images, the red area represents the target cells and the black area represents the interference cells.
[0088] like Figure 4 Figure a shows a WSI case of FL. After the annotator has marked all the tumor follicles with annotation lines, the following is obtained: Figure 4 The picture shown in Figure b.
[0089] Specifically, all annotation work is completed through the OpenHI annotation system. The system supports multi-user collaborative or crowdsourced annotation, enabling multiple experts or teams to work together, significantly improving the quality and efficiency of annotation. The OpenHI framework has pixel-level accurate annotation capabilities, which is essential for those tissue pathology image analysis that requires high precision. In addition, OpenHI supports standardized semantically rich annotations, which means that annotations are not just image tags, but also contain detailed information about tissues, diseases, etc., providing important value for research and education. The framework can effectively handle large histopathology image datasets and is suitable for large-scale research projects. The flexibility of its design allows support for annotation of whole slide images from different sources and tumor types.
[0090] According to the diagnostic guidelines, the area of each high-power field is set to 2 mm 2 , and in the annotation system used in the present invention, the size of an image block is 256x256 pixels. After conversion, it is determined that 10 high-power fields are equivalent to 40 regions of interest. Therefore, when performing external testing on each WSI case, 40 regions of interest are annotated for identifying and counting tumor cells, and histological grading is performed accordingly to improve the accuracy and effectiveness of grading.
[0091] 5. Image Cropping
[0092] The labeled data set is cropped and extracted to obtain several image blocks. Figure 2 As shown in Table 4, the image blocks are included in the training set, validation set, and internal test set in a ratio of 8:1:1. The specific number of image blocks in each group is shown in Table 4.
[0093] The model nuclear segmentation results are interpreted, and the Hover-NET model is used to realize the three-category recognition of cells at the image block level, specifically including centroblasts (positive) - C3, centroblasts (negative) - C2 and other cells (unclassified) - C1. To this end, the present invention divides the image blocks in the data set into the above three categories based on biological characteristics.
[0094] Table 4 Information on the number of graphic blocks in each group
[0095]
[0096] This allocation is intended to support subsequent neural network training and internal testing processes. During the internal testing phase, the positive cells identified by the model are compared with the manually annotated target cells for consistency. Only when the recognition consistency rate reaches or exceeds 95% will it enter the external testing phase. The external test set is mainly composed of cases that have been clearly diagnosed with follicular lymphoma.
[0097] like Figure 5 As shown in Figure a, it is a marked area of interest. Figure 5 The region of interest shown in Figure a is cropped to obtain Figure 5 Several image blocks shown in Figure b.
[0098] 6. Cell Nucleus Segmentation Model (Hover-NET)
[0099] Cell nucleus segmentation is a fundamental work for the study of cell characteristics in digital pathology analysis. Hover-Net is an innovative convolutional neural network designed for cell nucleus segmentation and classification of multi-tissue histological images. It has demonstrated excellent performance in the application of deep learning models, especially in the fields of medical image analysis and pathology image processing. This network not only plays a fundamental role in digital pathology analysis of H&E staining for the study of cell characteristics, but also plays a central role in the precise segmentation and classification of cell nuclei. This sophisticated analysis is crucial for understanding cellular heterogeneity in histological images, especially in identifying and classifying different cell types (such as tumor cells and immune cells). This sophisticated analysis helps pathologists gain a deep understanding of cellular heterogeneity in histological images, which in turn plays a key role in cancer diagnosis and treatment.
[0100] The architecture of Hover-Net combines the feature extraction capabilities of the pre-activated residual network with network branches specifically for segmentation and classification tasks, achieving efficient multi-task learning. This model is particularly suitable for processing digital slices. The "tile-based approach" adopted can effectively cope with the challenges of large-scale WSI and overcome the difficulties in processing large-size and high-complexity images. In addition, Hover-Net also adopts an innovative method to use the distance information from the nucleus pixel to the center of mass to improve the segmentation accuracy of overlapping nuclei, and optimizes the learning process for different task objectives through a composite loss function (including regression loss for segmentation tasks and cross entropy loss for classification tasks). It is worth noting that Hover-Net can perform cell type division while performing cell nucleus segmentation. This is achieved through multi-task learning, that is, the model simultaneously learns how to segment cell nuclei and identify their types.
[0101] In terms of performance evaluation, Hover-Net introduces a comprehensive evaluation metric called Pan-Perspective Quality (PQ). Performance on multiple datasets shows that it significantly outperforms other advanced nuclear segmentation methods. These technical innovations make Hover-Net a highly promising tool in the field of digital pathology, providing important support for oncology research.
[0102] 7. Hover-NET nuclear segmentation model training
[0103] The core feature of Hover-NET is to use the vertical and horizontal encoding information to deal with the problem of cell clustering, thereby achieving efficient nuclear segmentation. Nevertheless, in lymphoma sections, overlapping nuclei still pose a significant challenge because a large number of merged nuclei will affect the accuracy and sensitivity of nuclear size analysis.
[0104] The present invention found that the segmentation performance of the model was not ideal when processing WSIs with high nuclear overlap or thicker slices. In particular, in these complex WSIs, the Hover-NET nuclear segmentation model could not accurately segment each cell nucleus. To address this problem, the present invention customized and optimized the Hover-NET nuclear segmentation model.
[0105] The optimization process involves accurate manual annotation of each cell nucleus in three randomly selected regions of interest, and retraining the model using these annotation data to improve the accuracy of its nuclear segmentation. This optimization significantly improves the performance of the model, which can be verified by the improved Dice coefficient. The Dice coefficient is a widely used metric for evaluating the effectiveness of cell segmentation based on artificial intelligence algorithms. Before the optimization of the present invention, the Dice coefficient of the Hover-NET nuclear segmentation model was 0.78, and after optimization, this value was increased to 0.98. This significant improvement ensures accurate segmentation of cell nuclei in each ROI, providing a reliable basis for subsequent nuclear segmentation.
[0106] These classified images were then fed into a classifier to identify and distinguish the three types of cell images, e.g. Fig.10 In the output image of Figure a, the number of central mother cells is large, which is identified as high-level FL, and the low-level and Fig.10 In the output image of Figure b, there are more cells in the center, which is identified as low-level FL.
[0107] The present invention introduces a threshold-based classification mechanism. The classifier adopts a threshold-based discrimination method, in which the number of cells (denoted as x) is determined by the cell information output by the cell nucleus segmentation model. The classifier counts the number of cells and compares the number of centroblasts with a preset threshold (cutoff, denoted as C). When the number of centroblasts exceeds the threshold C, that is, x>C, the system automatically classifies the image as high-grade lymphoma tissue, otherwise it is classified as low-grade lymphoma tissue.
[0108] Specifically, when the image is classified as high-grade lymphoma tissue, the number 1 is output; when the image is classified as low-grade lymphoma tissue, the number 0 is output.
[0109] The features and performances of the present invention are further described in detail below in conjunction with application examples.
[0110] Application Example 1
[0111] When you select a picture such as Figure 6 When the pathological section image of a low-grade lymphoma is shown in Figure A, 40 regions of interest rich in diagnostic information are first screened in the pathological section image of the lymphoma; each region is 256×256 pixels;
[0112] Will be like Figure 6 The regions of interest shown in Figures B and C are input into the trained Hover-NET nuclear segmentation model, and the output is as follows Figure 6 The cell compartment image shown in Figure D;
[0113] Among them, the cell segmentation image can mark the areas of different cell types, the location of each cell and its classification information;
[0114] The cell segmentation image was input into the trained classifier, and the number of each cell type was output, including 1 centroblast, 26,556 centrocytes, and 3 non-tumor cells; the histological grading result was low grade.
[0115] Application Example 2
[0116] When you select a picture such as Figure 7 For the high-grade lymphoma pathology section image shown in Figure A, 40 typical regions are first screened in the lymphoma pathology section image, each region is 256×256 pixels;
[0117] Input the typical region (such as B and C in 7) into the trained Hover-NET nuclear segmentation model, and the output is as follows Figure 7 The cell compartment image shown in Figure D;
[0118] Among them, the cell segmentation image can mark the areas of different cell types, the location of each cell and its classification information;
[0119] The cell segmentation images were input into a trained classifier, and the number of each cell type was output, including 7183 centroblasts, 6973 centrocytes, and 667 non-tumor cells; the histological grading result was 3a.
[0120] Application Example 3
[0121] When you select a picture such as Figure 8 For the high-grade lymphoma pathology section image shown in Figure A, 40 typical regions are first screened in the lymphoma pathology section image, each region is 256×256 pixels;
[0122] Typical areas (such as Figure 8 B and C in Figure 1) are input into the trained Hover-NET nuclear segmentation model, and the output is as follows Figure 8 The cell compartment image shown in Figure D;
[0123] Among them, the cell segmentation image can mark the areas of different cell types, the location of each cell and its classification information;
[0124] The cell segmentation images were input into a trained classifier, and the number of each cell type was output, including 14,528 centroblasts, 28 centrocytes, and 281 non-tumor cells; the histological grading result was 3b.
[0125] The present invention also discloses a lymphoma histological grading system, comprising:
[0126] A screening module, used for screening multiple regions of interest rich in diagnostic information in lymphoma pathological slice images;
[0127] The nuclear segmentation module is used to input the region of interest into the trained Hover-NET nuclear segmentation model and output the cell segmentation image;
[0128] Among them, the cell segmentation image can mark the areas of different cell types, the location of each cell and its classification information;
[0129] The output module is used to input the cell compartment image into the classifier and output the histological grade results and the number of each cell type.
[0130] The lymphoma histological grading method of the present invention can be implemented in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention can be implemented in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
[0131] If the lymphoma histological grading method of the present invention is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. Computer-readable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals. Among them, the computer storage medium can be any available medium or data storage device that can be accessed by the computer, including but not limited to magnetic storage (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MO), etc.), optical storage (such as CD, DVD, BD, HVD, etc.), and semiconductor storage (such as ROM, EPROM, EEPROM, non-volatile memory (NANDFLASH), solid-state drive (SSD)), etc.
[0132] In an exemplary embodiment, a computer device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the lymphoma histological grading method when executing the computer program. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.
[0133] The present invention also discloses a computer program product, comprising a computer program / instruction, which implements the steps of the lymphoma histological grading method when executed by a processor.
[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for histological grading of lymphoma, characterized in that: The following steps are involved: S1. Screen multiple regions of interest rich in diagnostic information in lymphoma pathological slide images; S2, input the region of interest into the trained Hover-NET nuclear segmentation model and output the cell segmentation image; Among them, the cell segmentation image can mark the areas of different cell types, the location of each cell and its classification information; S3. Input the cell segmentation image into the classifier and output the histological grading results and the number of each cell type.
2. A method for histological grading of lymphoma according to claim 1, characterized in that: In S1, the tool used to screen multiple regions of interest rich in diagnostic information is the OpenHI annotation system.
3. A method for histological grading of lymphoma according to claim 1, characterized in that: In S2, the training process of the Hover-NET core segmentation model is: 2.
1. Obtain pathological section images of lymphoma with typical morphology and clinical diagnosis of grade 3b as the original image set of centroblasts; Acquire morphologically typical pathological section images of clinically diagnosed low-grade lymphoma, and some lightly stained low-grade lymphoma pathological section images, to form a central cell original image set; The original image set of centroblasts and the original image set of centroblasts constitute the original data set; 2.
2. Label the original data set, mark the effective tumor area and the interference cell area, and obtain the labeled data set; 2.
3. Crop the effective tumor area of each image in the labeled data set to obtain multiple image blocks; The image blocks obtained from the pathological section images of 2.4 and 3b grade lymphomas constitute the central mother cell dataset, and the image blocks obtained from the pathological section images of low-grade lymphomas constitute the central cell dataset. The two datasets are divided into training set, validation set and test set. The Hover-NET nuclear segmentation model is trained using the training set until the validation set is qualified and the test set meets the requirements. In this case, the Hover-NET nuclear segmentation model is qualified.
4. A method for histological grading of lymphoma according to claim 3, characterized in that: In 2.2, the tool used to annotate the original data set is the OpenHI annotation system.
5. A method for histological grading of lymphoma according to claim 4, characterized in that: In S3, the classifier can identify three types of cell image patches, which are divided into centroblasts, centroblasts, and other cells; The preset threshold in the classifier is the benchmark value of the number of centroblasts; After the classifier completes the statistics of the number of central lymphocytes, it compares it with the preset threshold; when the number of central lymphocytes exceeds the threshold, the image is classified as high-grade lymphoma tissue; otherwise, the image is classified as low-grade lymphoma tissue.
6. A method for histological grading of lymphoma according to claim 5, characterized in that: When the image is classified as high-grade lymphoma tissue, the output number is 1; When the image is classified as low-grade lymphoma tissue, the output is the number 0.
7. A lymphoma histological grading system, characterized in that: include: A screening module, used for screening multiple regions of interest rich in diagnostic information in lymphoma pathological slice images; The nuclear segmentation module is used to input the region of interest into the trained Hover-NET nuclear segmentation model and output the cell segmentation image; Among them, the cell segmentation image can mark the areas of different cell types, the location of each cell and its classification information; The output module is used to input the cell compartment image into the classifier and output the histological grade results and the number of each cell type.
8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the lymphoma histological grading method according to any one of claims 1 to 7 are implemented.
9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for histological grading of lymphoma according to any one of claims 1 to 7 are implemented.
10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the lymphoma histological grading method according to claim 1 are implemented.
Citation Information
Patent Citations
Pathological image classification method and system based on deep learning and machine learning
CN112215801A
Pathological image classification method and system and readable medium
CN113436191A
Method and device for interpreting lymphocytes
CN114299044A
Method for classifying cell nucleuses in full-view digital slice histopathological picture
CN114742758A
Glioma grading method, device and equipment based on histopathological image and medium
CN115082718A
