Deep learning-based kidney fibrosis grading method, electronic device, and storage medium
By constructing a two-stage weakly supervised learning framework and using deep learning methods to classify renal biopsy pathological images, the problem of poor consistency in visual assessment by pathologists is solved, and automatic and accurate diagnosis of the degree of renal fibrosis is achieved, supporting clinical diagnosis and treatment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2023-10-09
- Publication Date
- 2026-06-26
AI Technical Summary
In the current technology, the assessment of the degree of renal fibrosis relies on the visual assessment of pathologists, which has problems of poor consistency, misdiagnosis or missed diagnosis.
A two-stage weakly supervised learning framework was constructed to classify renal biopsy pathological images using deep learning methods. This included screening local slices that were strongly correlated with the whole slide digital image and performing splicing and decision aggregation. The ordinal regression loss function and the squared weighted kappa consistency coefficient were used as evaluation indicators.
It enables automatic and accurate diagnosis of the degree of renal fibrosis without the need for doctor intervention, providing objective diagnostic opinions and assisting pathologists in making treatment decisions.
Smart Images
Figure CN117408954B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a deep learning-based method for grading renal fibrosis, an electronic device, and a storage medium. Background Technology
[0002] Renal interstitial fibrosis is a hallmark of chronic kidney disease (CKD), and the degree of renal fibrosis is usually correlated with the severity of CKD. Currently, the assessment of the degree of renal interstitial fibrosis mainly relies on renal biopsy histopathological examination. Accurate diagnosis of renal fibrosis and implementation of treatment measures can slow the progression of CKD. However, due to poor consistency in the visual assessment of fibrosis among pathologists—even the same pathologist may make different assessments at different time points—and visual fatigue, misdiagnosis or missed diagnosis may occur. Summary of the Invention
[0003] This invention provides a deep learning-based method for grading renal fibrosis, an electronic device, and a storage medium. It constructs a two-stage deep learning framework to accurately identify and grade the degree of fibrosis in renal biopsy pathological images, thereby solving the problem of inaccurate manual grading.
[0004] A first aspect of this invention provides a deep learning-based method for grading renal fibrosis, comprising the following steps: acquiring a renal biopsy specimen to be graded, preparing the specimen, and scanning it to obtain a whole-slide imaging (WSI) digital image; segmenting the whole-slide imaging to obtain multiple whole-slide partial slice digital images; inputting the whole-slide partial slice digital images into a pre-trained grading model to obtain the renal fibrosis grading result of the renal biopsy specimen to be graded, wherein the grading model is a two-stage weakly supervised learning framework, in the first stage used to select multiple whole-slide partial slice digital images strongly correlated with the whole-slide imaging, and in the second stage used to stitch the multiple strongly correlated whole-slide partial slice digital images into a single digital image and feed it into a deep learning convolutional classification network for decision aggregation.
[0005] In one embodiment of the present invention, before inputting the digital image of the partial slice of the whole glass slide into the pre-trained hierarchical model, the method further includes:
[0006] Renal biopsy specimens from multiple patients with chronic kidney disease were obtained and slides were prepared. The slides were scanned to obtain digital images of the whole slides, and a dataset of renal fibrosis grading was constructed using the digital images of the whole slides.
[0007] The whole-slide digital images in the renal fibrosis grading dataset were labeled and divided into multiple fibrosis severity levels;
[0008] The whole glass slide digital image is segmented into multiple whole glass slide local slice digital images. A two-stage weakly supervised learning framework is trained using the multiple whole glass slide local slice digital images and their corresponding fiberization level labels to obtain the pre-trained hierarchical model.
[0009] In one embodiment of the present invention, obtaining a renal biopsy specimen to be graded, preparing a specimen, and scanning it to obtain a digital image of the entire slide include:
[0010] The renal biopsy specimens to be graded were fixed in neutral buffered formalin, dehydrated with graded alcohol, embedded in paraffin, sliced into 2-micrometer-thick sections, dewaxed and hydrated, stained with Sirius red, and scanned to obtain digital images of the whole slide.
[0011] In one embodiment of the present invention, before constructing a renal fibrosis grading dataset using the whole-slide digital image, the method further includes:
[0012] Remove poorly fixed tissue, inconsistent staining, tissue obscured by dirt, and unclear scanned digital images of the whole slide.
[0013] In one embodiment of the present invention, the digital image of the whole glass slide is segmented to obtain multiple digital images of partial slices of the whole glass slide, including:
[0014] The digital image of the whole glass slide is converted from RGB space to HSV space. Using the Saturation channel, a binary mask of the foreground of the tissue region is generated according to a preset threshold. The edges of the mask are smoothed by median filtering and morphological closure operation.
[0015] The digital image of the entire glass slide is downsampled, and multiple digital images of local slices of the entire glass slide are cropped from the segmentation contour at a preset size.
[0016] In one embodiment of the present invention, cropping multiple digital images of partial slices of a full-glass slide from a segmented contour at a preset size includes:
[0017] After the downsampled whole-slide digital image is cut into multiple non-overlapping whole-slide local slice digital images, the invalid background of the whole-slide local slice digital images is filtered out using the positive pixel counting method, and multiple whole-slide local slice digital images containing tissue are retained.
[0018] In one embodiment of the present invention, the first stage for filtering out multiple digital images of local slices of the whole slide that are strongly correlated with the digital image of the whole slide includes:
[0019] The Monte Carlo dropout method is used instead of the positive pixel counting method to filter out the full-slide digital images that are strongly correlated with the labels of the full-slide digital images. This eliminates the influence of label noise at the full-slide digital image level caused by weak labels in multi-instance learning. The prediction entropy is used to measure the uncertainty obtained by predicting each sample T times.
[0020] In one embodiment of the present invention, the second stage for stitching together multiple strongly correlated digital images of local slices of all-glass slides into a single digital image and feeding it into a deep learning convolutional classification network for decision aggregation includes:
[0021] For hierarchical tasks of deep learning convolutional integral networks, ordinal regression loss function is adopted, and squared weighted kappa consistency coefficient is used instead of accuracy as evaluation index.
[0022] A second aspect of the present invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to perform the deep learning-based renal fibrosis grading method as described in the above embodiments.
[0023] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to perform the deep learning-based grading method for renal fibrosis as described in the above embodiments.
[0024] This invention discloses a deep learning-based method for grading renal fibrosis, an electronic device, and a storage medium. By constructing a two-stage weakly supervised learning framework, it automatically grades the degree of fibrosis in renal biopsy pathological images. This grading method effectively applies artificial intelligence and image processing technologies to clinical medical diagnosis. It can automatically provide a diagnostic opinion on the degree of fibrosis in a patient's renal biopsy without physician intervention, and it possesses a certain degree of objectivity. It can assist nephrologists in diagnosis and provide guidance for subsequent treatment.
[0025] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0026] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:
[0027] Figure 1 A flowchart illustrating a deep learning-based grading method for renal fibrosis according to an embodiment of the present invention;
[0028] Figure 2A schematic diagram of the network structure of a two-stage weakly supervised learning framework constructed according to an embodiment of the present invention;
[0029] Figure 3 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of the present invention. Detailed Implementation
[0030] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0031] Figure 1 This is a flowchart of a deep learning-based grading method for renal fibrosis provided according to an embodiment of the present invention.
[0032] like Figure 1 As shown, this deep learning-based method for grading renal fibrosis includes the following steps:
[0033] In step S101, a renal biopsy specimen to be graded is obtained and prepared, and a digital image of the whole slide is obtained by scanning. The digital image of the whole slide is then segmented to obtain multiple digital images of local sections of the whole slide.
[0034] In one embodiment of the present invention, obtaining a renal biopsy specimen to be graded, preparing a specimen, and scanning it to obtain a digital image of the entire slide include:
[0035] The renal biopsy specimens to be graded were fixed in neutral buffered formalin, dehydrated with graded alcohol, embedded in paraffin, sliced into 2-micrometer-thick sections, dewaxed and hydrated, stained with Sirius red, and scanned to obtain digital images of the whole slide.
[0036] For patients with chronic kidney disease, biopsy samples should be fixed in a timely manner, with high-quality tissue sections free of scalpel marks and folds, large tissue blocks, and consistent conditions such as staining reagents, time, temperature, and humidity. The scanning equipment should also have high resolution.
[0037] In one embodiment of the present invention, the digital image of the whole glass slide is segmented to obtain multiple digital images of partial slices of the whole glass slide, including:
[0038] The digital image of the whole glass slide is converted from RGB space to HSV space. Using the Saturation channel, a binary mask of the foreground of the tissue region is generated according to a preset threshold. The edges of the mask are smoothed by median filtering and morphological closure operation.
[0039] The digital image of the whole glass slide is downsampled, and multiple digital images of local slices of the whole glass slide are cropped from the segmentation contour at a preset size.
[0040] The original digital image file of the whole glass slide has a large pixel size, which cannot be directly used for training of the neural network. Therefore, the digital image of the whole glass slide is divided into hundreds or thousands of local slices (hereinafter referred to as "patches") as network input.
[0041] Specifically, the details of image preprocessing, specifically segmenting the digital image of a whole slide into patches, are as follows: First, for each digital image of a whole slide, it is converted from RGB space to HSV space. Using the saturation channel, a binary mask for the foreground of the tissue region is generated according to a certain threshold. The edges of the mask are smoothed by median filtering and morphological closure operations. Second, the original image (40 times resolution) is downsampled to 10 times resolution (the resolution commonly used by physicians for analysis), and patches are cropped from the segmentation contour at a size of 224×224.
[0042] In one embodiment of the present invention, cropping multiple digital images of partial slices of a full-glass slide from a segmented contour at a preset size includes:
[0043] After the downsampled whole-slide digital image is segmented into multiple non-overlapping whole-slide local slice digital images, the invalid background of the whole-slide local slice digital images is filtered out using the positive pixel counting method, and multiple whole-slide local slice digital images containing tissue are retained.
[0044] The digital image of the whole slide was cut into hundreds of non-overlapping patches at a 10x magnification tiling field of view. Patches with invalid background were filtered out using a positive pixel counting method, leaving the patches containing the tissue.
[0045] In step S102, the digital images of partial slices of the whole slide are input into a pre-trained grading model to obtain the grading results of renal fibrosis of the renal biopsy specimen to be graded. The grading model is a two-stage weakly supervised learning framework. In the first stage, it is used to select multiple digital images of partial slices of the whole slide that are strongly correlated with the digital images of the whole slide. In the second stage, it is used to stitch the multiple strongly correlated digital images of partial slices of the whole slide into a single digital image and feed it into a deep learning convolutional classification network for decision aggregation.
[0046] In one embodiment of the present invention, before inputting the digital image of a partial slice of the whole glass slide into the pre-trained hierarchical model, the method further includes:
[0047] Renal biopsy specimens from multiple patients with chronic kidney disease were obtained, slides were prepared, and the slides were scanned to obtain digital images of the whole slide. A dataset of renal fibrosis grading was constructed using the digital images of the whole slide.
[0048] The whole-slide digital images in the renal fibrosis grading dataset were labeled and divided into multiple fibrosis levels.
[0049] The digital image of the whole glass slide is segmented into multiple digital images of local slices of the whole glass slide. A two-stage weakly supervised learning framework is trained using multiple digital images of local slices of the whole glass slide and their corresponding fiberization level labels to obtain a pre-trained hierarchical model.
[0050] To train the grading model, a training dataset was first constructed. Renal biopsy specimens from patients with chronic kidney disease were collected, fixed in neutral-buffered formalin, dehydrated with gradient ethanol, embedded in paraffin, sliced into 2-micrometer-thick sections, dewaxed and hydrated, stained with Sirius red, and scanned into whole-slide digital images. To improve training effectiveness, a dataset for grading renal fibrosis was constructed by selecting suitable sections, and whole-slide digital images corresponding to sections with poor tissue fixation, inconsistent staining, tissue obscured by dirt, or unclear scans were removed.
[0051] The degree of fibrosis was assessed using digital images of whole slides using image processing software, and two pathologists visually evaluated the renal biopsy slides, classifying the degree of fibrosis into four levels: very mild, mild, moderate, and severe. Slide samples with consistent assessments by the two pathologists were used as a dataset for training and testing the model.
[0052] The image processing software used was Image Scope. The two physicians were unaware of the clinical information of the slides beforehand, and the two pathologists conducted the evaluation without communicating with each other. The samples used to train the model had a balanced degree of fibrosis.
[0053] like Figure 2 As shown, a two-stage weakly supervised learning framework is established. In the first stage, patches that are strongly correlated with the labels of the whole slide digital images are selected. In the second stage, the strongly correlated patches are stitched together into a large image, which is then fed into a classification network model for decision aggregation to obtain the diagnostic evaluation results at the whole slide digital image level.
[0054] Specifically, the graded dataset includes 465 kidney biopsy pathology images, and a total of 367 images in the training and validation sets (using five-fold cross-validation).
[0055] In one embodiment of the present invention, the first stage for filtering out multiple digital images of local slices of the whole slide that are strongly correlated with the digital image of the whole slide includes:
[0056] By using MC Dropout instead of positive pixel counting, we select full-slide local slice digital images that are strongly correlated with the labels of the full-slide digital images. This eliminates the influence of label noise at the full-slide local slice digital image level caused by weak labels in multi-instance learning. We use prediction entropy to measure the uncertainty obtained by predicting each sample T times.
[0057] In the patch selection phase, Monte Carlo Dropout (MC Dropout) is used instead of positive pixel counting to select patches that are strongly correlated with the labels of the whole slide digital image. This eliminates the patch-level label noise caused by weak labels in multi-instance learning. Prediction entropy is used to measure the uncertainty obtained from predicting each sample T times. The lower the uncertainty, the more confident the model is in predicting that sample, meaning it is more correlated with the labels of the entire original image.
[0058] In one embodiment of the present invention, the second stage for stitching together multiple strongly correlated digital images of local slices of all-glass slides into a single digital image and feeding it into a deep learning convolutional classification network for decision aggregation includes:
[0059] For hierarchical tasks of deep learning convolutional integral networks, ordinal regression loss function is adopted, and squared weighted kappa consistency coefficient is used instead of accuracy as evaluation index.
[0060] The selected patches are stitched together into a large image, which is then fed into a deep learning convolutional classification network for decision aggregation to obtain the classification result of the entire slide. For the classification task, an ordinal regression loss function is used instead of a regular loss function, and a squared weighted kappa consistency coefficient is used instead of accuracy as the evaluation metric, thus taking into account the characteristics of both classification and regression tasks.
[0061] Specifically, this paper proposes an automated grading framework for super-resolution renal biopsy slide fibrosis based on approximate Bayesian inference. This framework consists of two stages. The first stage (patch selection stage) utilizes Monte Carlo dropout to approximate Bayesian inference, estimating uncertainty and selecting patches with low uncertainty, i.e., those highly correlated with the WSI label. The second stage (decision aggregation stage) aggregates the top N most relevant patches to the WSI label. 2 Each patch is extracted, stitched together into a large N×M image, and then fed into a classification network for decision aggregation to obtain the WSI fiber grading assessment results.
[0062] The first stage, the patch selection stage, involves segmenting the WSI (Wide Image Segmentation) into hundreds of non-overlapping patches at a 10x magnification tiling. Invalid background patches are filtered out using positive pixel counting, leaving patches containing organization. Each patch is assigned a WSI-level label as a supervision label for the classification network. The lightweight model EifficientNet was chosen to save time required for subsequent variational inference. Subsequently, to achieve uncertainty estimation, we slightly modified the last pooling layer of EifficNet by adding two fully connected layers. Each fully connected layer is followed by a ReLU non-linear activation layer and a Dropout layer, with a dropout rate set to 0.5. For T predictions, the probability can be approximated using Monte Carlo integration, as shown below:
[0063]
[0064] in, It follows a Dropout distribution.
[0065] Prediction entropy is used to measure the uncertainty of each sample.
[0066]
[0067]
[0068] in, This represents the softmax probability of class c in the t-th prediction. The samples are divided into a total of C classes. The prediction entropy ranges from 0 to 1; the closer it is to 1, the greater the uncertainty and the lower the confidence in predicting the sample.
[0069] Therefore, based on the uncertainty estimate, the top N most relevant to the WSI label can be selected. 2 Each patch was selected and incorporated into the second stage.
[0070] In the first stage, N is obtained through MC dropout and variational inference. 2 These are low-uncertainty patches. These patches are concatenated to obtain an N×M concatmap. Before feeding the concatmap into the classification network, several techniques are used for further data augmentation, including shiftaugmentation, spraltile, random sorting, random rotation, and random whitening. Additionally, there are data augmentation strategies for the global graph, including transpose, horizontal flip, vertical flip, random rotation, random translation, and non-rigid transformations.
[0071] After data augmentation, the concatmap is fed into the classification network, and finally passes through a fully connected layer and a softmax layer to obtain the WSI fiber classification evaluation results.
[0072] Using ordinal regression loss function L OR This approach balances the characteristics of both classification and regression tasks. Considering a classification network based on class order, x can be encoded as t = (1, ..., 1, 1, 0, ..., 0), setting the first k elements to 1 and the rest to 0. This transforms an ordinal regression task into a multi-label classification task, and then uses the cross-entropy loss function L... CE To calculate.
[0073]
[0074] Pathological images often exhibit class imbalance, with negative cases (or mild cases) far outnumbering positive cases (or severe cases). Therefore, a weighted loss function can be used as follows:
[0075]
[0076] Among them, the weight term Balance the loss L for each class c c N is the total number of data sets. c The amount of data belongs to class c, and k is a factor that controls the balance of importance between classes. In this paper, k = 2.
[0077] The deep learning-based grading method for renal fibrosis proposed in this invention automatically grades the degree of fibrosis in renal biopsy pathological images by constructing a two-stage weakly supervised learning framework. This grading method effectively applies artificial intelligence and image processing technologies to clinical medical diagnosis, automatically providing a diagnostic opinion on the degree of renal biopsy fibrosis without physician intervention. It also possesses a degree of objectivity, assisting nephrologists in diagnosis and providing guidance for subsequent treatment.
[0078] Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. The electronic device may include:
[0079] The memory 301, the processor 302, and the computer program stored on the memory 301 and capable of running on the processor 302.
[0080] When the processor 302 executes the program, it implements the deep learning-based grading method for renal fibrosis provided in the above embodiments.
[0081] Furthermore, the vehicle also includes:
[0082] Communication interface 303 is used for communication between memory 301 and processor 302.
[0083] The memory 301 is used to store computer programs that can run on the processor 302.
[0084] The memory 301 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.
[0085] If the memory 301, processor 302, and communication interface 303 are implemented independently, then the communication interface 303, memory 301, and processor 302 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (ELSA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.
[0086] Optionally, in a specific implementation, if the memory 301, processor 302, and communication interface 303 are integrated on a single chip, then the memory 301, processor 302, and communication interface 303 can communicate with each other through an internal interface.
[0087] Processor 302 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.
[0088] This embodiment also provides a computer-readable storage medium storing a computer program thereon, characterized in that the program, when executed by a processor, implements the deep learning-based renal fibrosis grading method described above.
[0089] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0090] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0091] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.
[0092] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0093] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.
Claims
1. A deep learning-based method for grading renal fibrosis, characterized in that, Includes the following steps: Obtain the renal biopsy specimen to be graded, prepare the specimen, and scan to obtain a digital image of the whole slide. Segment the digital image of the whole slide to obtain multiple digital images of partial sections of the whole slide. The digital images of partial sections of the whole slide are input into a pre-trained grading model to obtain the grading result of renal fibrosis of the renal biopsy specimen to be graded. The grading model is a two-stage weakly supervised learning framework. In the first stage, it is used to select multiple digital images of partial sections of the whole slide that are strongly correlated with the digital images of the whole slide. In the second stage, it is used to stitch the multiple strongly correlated digital images of partial sections of the whole slide into a single digital image and feed it into a deep learning convolutional classification network for decision aggregation. In the first stage, multiple full-glass slide local slice digital images that are strongly correlated with the full-glass slide digital image are selected, including: using the Monte Carlo dropout method instead of the positive pixel counting method to select full-glass slide local slice digital images that are strongly correlated with the labels of the full-glass slide digital images, eliminating the influence of label noise at the full-glass slide local slice digital image level caused by weak labels in multi-instance learning, and using prediction entropy to measure the uncertainty obtained by predicting each sample T times. In the second stage, multiple strongly correlated digital images of local slices of all-glass slides are stitched together into a single digital image and fed into a deep learning convolutional class network for decision aggregation. This includes: for the hierarchical task of the deep learning convolutional class network, an ordinal regression loss function is used, and the squared weighted kappa consistency coefficient is used instead of accuracy as the evaluation index.
2. The method according to claim 1, characterized in that, Before inputting the digital image of the partial slice of the whole glass slide into the pre-trained hierarchical model, the process also includes: Renal biopsy specimens from multiple patients with chronic kidney disease were obtained and slides were prepared. The slides were scanned to obtain digital images of the whole slides, and a dataset of renal fibrosis grading was constructed using the digital images of the whole slides. The whole-slide digital images in the renal fibrosis grading dataset were labeled and divided into multiple fibrosis severity levels; The whole glass slide digital image is segmented into multiple whole glass slide local slice digital images. A two-stage weakly supervised learning framework is trained using the multiple whole glass slide local slice digital images and their corresponding fiberization level labels to obtain the pre-trained hierarchical model.
3. The method according to claim 1, characterized in that, Obtain the renal biopsy specimen to be graded, prepare the specimen, and scan it to obtain a digital image of the entire slide, including: The renal biopsy specimens to be graded were fixed in neutral buffered formalin, dehydrated with graded alcohol, embedded in paraffin, sliced into 2-micrometer-thick sections, dewaxed and hydrated, stained with Sirius red, and scanned to obtain digital images of the whole slide.
4. The method according to claim 1, characterized in that, Before constructing a kidney fibrosis grading dataset using the whole-slide digital images, the following steps are also included: Remove poorly fixed tissue, inconsistent staining, tissue obscured by dirt, and unclear scanned digital images of the whole slide.
5. The method according to claim 1, characterized in that, The digital image of the entire glass slide is segmented to obtain multiple digital images of partial slices of the entire glass slide, including: The digital image of the whole glass slide is converted from RGB space to HSV space. Using the Saturation channel, a binary mask of the foreground of the tissue region is generated according to a preset threshold. The edges of the mask are smoothed by median filtering and morphological closure operation. The digital image of the entire glass slide is downsampled, and multiple digital images of local slices of the entire glass slide are cropped from the segmentation contour at a preset size.
6. The method according to claim 5, characterized in that, Digital images of multiple local slices of a full glass slide, cropped from the segmented contour at a preset size, including: After the downsampled whole-slide digital image is cut into multiple non-overlapping whole-slide local slice digital images, the invalid background of the whole-slide local slice digital images is filtered out using the positive pixel counting method, and multiple whole-slide local slice digital images containing tissue are retained.
7. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the deep learning-based grading method for renal fibrosis as described in any one of claims 1-6.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the deep learning-based grading method for renal fibrosis as described in any one of claims 1-6.