Hepatocellular carcinoma pathological micro-necrosis recognition system and method based on federal continuous learning

Abstract

The invention discloses a hepatocellular carcinoma pathological micro-necrosis recognition system and method based on federal continuous learning. The method comprises the following steps: firstly, obtaining time sequence image data of a multi-center full-size hepatocellular carcinoma tumor pathological section; performing foreground segmentation on the image data, removing a non-tumor background region, and generating a classification task-oriented standardized image block group; then, after each round of training is completed, the local model weight of each center is uploaded to a server, the server calculates the global weight according to the data scale of each center by adopting a weighted average strategy based on the data sample size, and the global weight is distributed to all participating centers for model iteration; a timing sequence sensitive regularization constraint is designed, and the weight updating strength of local training of each center is dynamically adjusted; and finally, comprehensively calculating the cross entropy loss and the proportion label fitting loss of the image block classification task, and the regular term loss in the model iteration process.

Description

Technical Field

[0001] The present invention relates to the technical field of pathological image recognition, and particularly to a hepatocellular carcinoma pathological micro-necrosis recognition system and method based on federated continuous learning. Background Art

[0002] Histopathological image analysis is a key step in diagnosing various cancers. The process begins with a pathologist obtaining a representative tissue sample from a patient through a biopsy. These samples then undergo a series of preparation steps, including fixation, dehydration, clearing, impregnation, embedding, and sectioning, ultimately forming a pathological section. After the section is stained and mounted, it is ready for the pathologist to observe carefully under a microscope. In this process, hematoxylin and eosin (H&E staining) are the most commonly used stains in histopathology. These dyes can specifically interact with specific structures or components within biological tissues, causing different structures in the tissue section to exhibit different colors. This color difference helps to reveal the morphology, structure, and distribution of cells, providing rich histological information for pathologists and thus playing an indispensable role in clinical diagnosis. Through these detailed visual information, pathologists can accurately diagnose diseases and formulate appropriate treatment plans for patients. However, this examination consumes a large amount of time, and the results are highly subjective. Several previous diagnoses in histopathological grading have shown poor consistency among pathologists. In addition, pathologists may miss small cancerous areas.

[0003] Against the backdrop of the digital revolution, digital slide scanners convert pathological sections into digital pathological images (WholeSlide Images, WSI), facilitating subsequent applications in preservation, teaching, remote reading, etc. In recent years, digital pathology and computational pathology have developed rapidly with the help of scanning devices, computing servers, and deep learning algorithms. Research on WSI of various diseases has achieved good results in cancer diagnosis, subtype classification, and prognostic stratification. Computer-aided diagnosis systems using deep learning algorithms can not only alleviate the drawbacks of pathologist diagnosis but also improve diagnosis by screening out obvious benign pathological sections and providing quantitative characterization of suspicious areas.

[0004] Therefore, many computer-aided pathological diagnosis systems are also based on deep learning models trained on different tasks and applied to clinical practice. However, the vast majority of computer-aided pathological diagnosis systems often train deep learning models based on data from a single medical center. Because in the real-world clinical environment, there are significant heterogeneities in the feature distributions of patients in different medical centers. In addition, since different medical institutions use different preparation procedures, stains, and scanners in the process of preparing digital pathological images, resulting in different appearances of tissue sections, it will also affect the robustness and generalization ability of deep learning models. At the same time, due to the huge size of a single WSI, the WSI dataset of a medical institution is often on the order of hundreds of GB or even TB. There are many practical problems such as storage and data transmission in aggregating the WSI data of many medical institutions for model training. The high privacy of medical data also makes centralized training more difficult. In addition, since the distribution of the patient population also changes dynamically over time, in order to meet the actual needs of computer-aided pathological diagnosis systems in clinical practice, it is necessary to regularly and continuously conduct local evaluations of the model to further ensure the clinical reliability of the model. Therefore, designing a federated continuous learning system that can effectively and accurately reflect clinical needs and can utilize data from multiple medical centers for learning under the premise of privacy protection, and can always maintain the best performance in each medical center, is of great significance to the clinical practice of computer-aided pathological diagnosis systems and can improve the generalization and robustness of computer-aided pathological diagnosis systems to a certain extent.

[0005] Currently, to solve this problem, many studies have attempted to use federated learning techniques to improve the generalization ability of computer-aided pathological diagnosis systems across multiple centers. Among them, the technical solution closest to the one claimed in this patent is as follows: ① Model optimization based on improved federated aggregation. This method addresses data heterogeneity by improving the model aggregation mechanism in federated learning. Federated Proximal (FedProx) (LI T, SAHU A K, ZAHEER M, et al. Federated Optimization in Heterogeneous Networks[J]. Proceedings of Machine Learning and Systems, 2020, 2: 429-450.) adds a "proximal term" to the local objective function, which acts as a regularizer and penalizes large deviations from the global model. Simply put, it restricts local updates to make them closer to the global model. In another method, Stochastic Controlled Averaging in Federated Learning (SCAFFOLD) (KARIMIREDDY S P, KALE S, MOHRI M, et al. SCAFFOLD: Stochastic Controlled Averaging for Federated Learning[C / OL] / / Proceedings of the 37th International Conference on Machine Learning. PMLR, 2020: 5132-5143[2024-11-29]. https: / / proceedings.mlr.press / v119 / karimireddy20a.html.) uses local control variables at each client to estimate the global update direction and adjusts local updates to reduce the bias introduced by heterogeneous data. ② Model alignment mechanism under data silos.This technical solution addresses the distribution differences of multi-center medical data and constrains the local model. S.M. Hosseini et al. proposed a proportionally fair federated learning (Prop-FFL) (HOSSEINI S M, SIKAROUDI M, BABAIE M, et al. Proportionally Fair Hospital Collaborations in Federated Learning of Histopathology Images[J / OL]. IEEE Transactions on Medical Imaging, 2023, 42(7): 1982-1995. DOI:10.1109 / TMI.2023.3234450.) aiming to overcome the bias in client considerations by introducing a second loss objective that rewards similar training losses for all clients while also considering the proportion of training samples for each client. Clients with fewer samples receive lower weights. S. Albarqouni et al. proposed the SiloBN method (Siloed Federated Learning for Multi-centric Histopathology Datasets | SpringerLink[EB / OL]. [2024-11-29]. https: / / link.springer.com / chapter / 10.1007 / 978-3-030-60548-3_13.), which keeps the batch normalization (BN) parameters of each client private and excludes them from model aggregation. This allows the global model to adapt to each local dataset but requires calculating the BN statistics for any unseen dataset during inference.

[0006] Similar existing technologies as ① improve the model generalization ability through the improvement of the federated learning aggregation method, but it is difficult to adapt to the dynamic evolution characteristics of multi-center data, and there are the following problems:

[0007] 1. Mismatch between static regularization and dynamic data: FedProx forces the local model to align with the global model through fixed parameter constraints, but it cannot adapt to the time-evolution characteristics of the multi-center data distribution. When the data distribution shifts over time, the fixed constraints are likely to lead to a cyclic contradiction of "underfitting - divergence" in model updates.

[0008] 2. Incremental updates forget historical knowledge: Traditional federated aggregation only focuses on cross-center spatial generalization and does not design a continuous learning mechanism in the time dimension, resulting in the loss of key pathological features learned in the past due to catastrophic forgetting during incremental updates.

[0009] 3. Global aggregation ignores local temporal evolution: SCAFFOLD corrects the local update direction by controlling variables, but its global update direction is only based on the current data distribution and does not model the dynamic evolution path of the local data at each center, resulting in model performance degradation during long-term deployment.

[0010] Similar existing technologies adopt a static distribution alignment strategy to improve the cross-domain generalization ability of the model for heterogeneous target populations, but they rely too much on the assumption of a fixed data distribution and have the following problems:

[0011] 1. Conflict between local constraints and global generalization: Prop-FFL balances the losses of each center by weighting the sample ratio, but the static weight allocation cannot adapt to the dynamic data distribution shift. For example, when a rare case type is added to a single center, its sample ratio weight is underestimated, resulting in limited generalization ability of the model for new scenarios.

[0012] 2. Insufficient task adaptability: Existing alignment mechanisms only optimize the consistency of the parameter distribution between model layers and do not design a cross-center feature alignment strategy for specific diagnostic tasks, resulting in the decoupling of the standardized feature space from the diagnostic task objective and limited recognition accuracy. Summary of the Invention

[0013] The object of the present invention is to propose a multi-center hepatocellular carcinoma tumor pathological micro-necrosis recognition system and method based on federated continuous learning in view of the deficiencies of the existing technologies.

[0014] The object of the present invention is achieved through the following technical solutions: In the first aspect, the present invention provides a multi-center hepatocellular carcinoma tumor pathological micro-necrosis recognition system based on federated continuous learning, and the system includes:

[0015] A dataset construction module, configured to obtain the temporal image data of the full-size hepatocellular carcinoma tumor pathological sections of each center;

[0016] A data preprocessing module, configured to perform foreground segmentation on the image data, remove the non-tumor background area, and generate a standardized image block group for the classification task;

[0017] A federated aggregation module, configured to upload the local model weights of each center to the server after each round of training is completed, and the server calculates the global weights using a weighted average strategy based on the data sample size according to the data scale of each center, and distributes them to all participating centers for model iteration;

[0018] A time perception module, configured to design a time-sensitive regularization constraint to dynamically adjust the weight update intensity of local training at each center;

[0019] A loss calculation module is used to comprehensively calculate the cross - entropy loss and the proportional label fitting loss of the image patch classification task, as well as the regularization loss during the model iteration process.

[0020] Furthermore, the image data is a full - size digital scan image of a tumor tissue pathological section obtained by hematoxylin - eosin staining, and a proportional label vector is constructed for the pathological image.

[0021] Furthermore, the data pre - processing module calculates the thumbnail size of the pathological image at different magnifications according to the set target image patch overlap rate, and uses the difference distribution characteristics of the RGB channels and grayscale values of the thumbnail to segment the foreground of the tumor tissue, remove the background of the tumor tissue, perform pixel - level segmentation on the thumbnail, where each pixel point corresponds to an image patch, and the pixel points are retained and excluded according to the segmentation threshold of the Otsu threshold segmentation method.

[0022] Furthermore, for the first - batch data set of each center, the federated aggregation module performs noise - robust training on the local server and optimizes the local model parameters through a fuzzy proportional label learning strategy.

[0023] Furthermore, the data sets of subsequent batches of each center are regarded as an incremental task flow corresponding to the pathological data distribution in a specific time period. Based on the feature similarity of the new and old data distributions, the retention intensity of historical parameters is automatically enhanced:

[0024]

[0025] where, (T - t) represents the difference between the current batch year and the historical year, and a is the decay coefficient, ensuring that the forgetting rate of the model for early data decreases year by year during model update.

[0026] Furthermore, the cross - entropy loss of the image patch classification task includes the losses at different magnifications of the image patches, and through random up - sampling or random down - sampling, the training image patches in one training batch all come from the same pathological section.

[0027] Furthermore, the regularization loss is specifically as follows:

[0028]

[0029] where, represents the optimal model parameters in the historical training stage, which are used as anchor points to constrain the update direction of the current parameters to avoid the degradation of the recognition ability for historical data distributions, adjusts the weights of the new and old data distributions, and its value decays with the number of training rounds to gradually adapt to the evolving characteristics of new data; It is a quantization parameter of the Fisher Information Matrix (FIM). The importance weight in the historical model. The higher the importance of a parameter, the greater the retention strength in the regularization constraint, thus preferentially protecting key pathological features.

[0030] In a second aspect, the present invention also discloses a method for identifying pathological micro - necrosis of hepatocellular carcinoma based on federated continuous learning, which includes the following steps:

[0031] (1) Obtain the time - series image data of full - size pathological sections of hepatocellular carcinoma tumors from each center;

[0032] (2) Perform foreground segmentation on the image data, remove the non - tumor background area, and generate a standardized image block group for the classification task;

[0033] (3) After each round of training, upload the local model weights of each center to the server. The server calculates the global weights using a weighted average strategy based on the data sample size according to the data scale of each center, and distributes them to all participating centers for model iteration;

[0034] (4) Design a time - series - sensitive regularization constraint to dynamically adjust the weight update strength of local training in each center;

[0035] (5) Comprehensively calculate the cross - entropy loss and the proportion label fitting loss of the image block classification task, as well as the regularization term loss during the model iteration process.

[0036] In a third aspect, the present invention also discloses a device for identifying pathological micro - necrosis of hepatocellular carcinoma based on federated continuous learning, including a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the method for identifying pathological micro - necrosis of hepatocellular carcinoma based on federated continuous learning as described above.

[0037] In a fourth aspect, the present invention also discloses a computer - readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the method for identifying pathological micro - necrosis of hepatocellular carcinoma based on federated continuous learning as described above.

[0038] Advantages of the present invention:

[0039] (1) The present invention introduces a dynamic regularization constraint to replace the fixed constraint, ensuring the strong generalization ability of the model in the clinical environment when the data distribution drifts over time.

[0040] (2) The present invention uses a training strategy of cyclic local validation, combined with the optimization of the joint loss function to achieve task - driven cross - center and cross - time feature alignment, which helps to realize the training and deployment of the model in the real - world clinical scenario.

[0041] (3) The present invention breaks through the limitation of domain shift on micro-necrosis recognition through a dynamic weighted federated aggregation strategy.

[0042] (4) The present invention constructs a model update strategy with time-aware constraints under incremental data to solve the contradiction between catastrophic forgetting of historical tasks and adaptability to dynamic data distribution.

[0043] (5) The present invention designs a multi-modal attention fusion module, combines morphological prior knowledge with deep feature expression, and establishes a cross-center robust micro-necrosis quantification and evaluation system. Brief Description of the Drawings

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0045] Figure 1 It is the system structure diagram of the present invention.

[0046] Figure 2 It is the federated continuous learning architecture diagram of the present invention.

[0047] Figure 3 It is a schematic structural diagram of a device for identifying hepatocellular carcinoma pathological micro-necrosis based on federated continuous learning. Detailed Embodiments

[0048] In order to make the purpose, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described below in conjunction with the drawings and implementation cases. It should be understood that the specific implementation cases described herein are only used to explain the present invention and are not used to limit the present invention.

[0049] As Figure 1 shown, a multi-center hepatocellular carcinoma tumor pathological micro-necrosis recognition system based on federated continuous learning provided by the present invention mainly includes a dataset construction module, a data preprocessing module, a federated aggregation module, a time perception module, and a loss calculation module.

[0050] Among them, the dataset module constructs a structured dataset containing full-size hepatocellular carcinoma tumor pathological sections based on multi-center dynamic time-series data; the data preprocessing module adopts a multi-scale foreground segmentation strategy, combines Otsu threshold segmentation and image patch recombination technology to remove non-tumor background regions, generates a standardized image patch group for classification tasks, improves data quality and model training efficiency, and supports the standardized expression of cross-center heterogeneous data; the federated aggregation module realizes the collaborative optimization of multi-center models through a robust training and data volume weighted synchronous aggregation mechanism, alleviates the impact of heterogeneous data distribution offset on global generalization performance; the time-aware module introduces dynamic regularization constraints to adaptively adjust the weights of new and old data, suppresses catastrophic forgetting and enhances the model's continuous adaptation ability to dynamic data distribution; the loss calculation module fuses cross-entropy loss, proportional label fitting loss and time-aware regularization terms, and realizes the balanced improvement of the recognition accuracy of micro-necrotic regions and the retention of historical knowledge through a multi-magnification joint optimization strategy. The specific implementation process of each module is as follows:

[0051] 1. Dataset construction module:

[0052] For a single client data center, the obtained dataset is denoted as B, which contains a total of n full-size digital scans (WholeSlide Image, WSI) of tumor tissue pathological sections obtained by hematoxylin-eosin (HE) staining, that is , and the classification task target has a total of C categories. For the i-th pathological image , the constructed proportional label vector is: ;

[0053] Among them, is the proportional threshold of the k-th category in the i-th sample image, and satisfies . The proportional threshold refers to the lowest value of the proportional range during manual rough estimation. If there are a total of centers, then there is a total dataset , where the dataset of the th center can be divided into sub-datasets , represents the j-th batch of collections. After the first batch of dataset training, the model will be dynamically updated according to the new batch of datasets.

[0054] 2. Data preprocessing module:

[0055] Adopt a multi - center pathological data standardization pre - processing process. Based on foreground segmentation and proportional label vectors, standardize the annotation of HE - stained pathological sections, break through the limitations of multi - center data heterogeneity, and reduce the interference of non - tumor regions. The first step of pre - processing is foreground segmentation. At different magnifications of pathological images, calculate the thumbnail size of the pathological image according to the set target image patch overlap rate, and use the difference distribution characteristics of the RGB channels and grayscale values of the thumbnail to segment the foreground of tumor tissue, remove the background of tumor tissue, perform pixel - level segmentation on the thumbnail. Each pixel point corresponds to an image patch. For pixel points, retain and eliminate them according to the segmentation threshold of the Otsu threshold segmentation method. For a single data set, finally obtain the total data set after removing the background , including sample packs, and the i - th sample pack contains groups of example image patches, denoted as , indicating the groups of image patches after foreground segmentation at different magnifications.

[0056] 3. Federal aggregation module:

[0057] Based on the image patch classification problem, the present invention adopts a federal synchronous parameter aggregation mechanism to achieve cross - center collaborative training. Based on the training architecture of federal continuous learning and the local loop verification strategy, it avoids the offset of patient pathological features due to time changes and meets the requirements of real - world clinical scenarios. The specific process is as follows:

[0058] (1) Local model training: For the first - batch data sets of K centers , perform noise - robust training (Noise - Robust Training, NRT) on the local server, and optimize the local model parameters through the fuzzy proportional label learning strategy;

[0059] (2) Synchronous parameter aggregation: After each round of training, each center synchronously uploads the local model weights to the central server. The server calculates the global weights (i.e., the total number of image patches of the i - th center) according to the data scale of each center, and adopts a weighted average strategy based on the data sample size to calculate the global weights where, is the full sample size, and this strategy can effectively balance the impact of heterogeneous data distribution on the global model.

[0060] (3) Global model distribution and iterative update: The server distributes to all participating centers, and each center starts the next round of local training based on the updated global model, forming a closed - loop optimization process.

[0061] 4. Time - aware module:

[0062] Through the federated continuous learning framework, design a time-sensitive regularization constraint to dynamically adjust the weight update intensity of local training at each center. Introduce a time-aware dynamic regularization coefficient in distributed incremental training, and combine the data volume weighting strategy to achieve adaptive allocation of cross-center model weights. After the aggregated training of the first batch of models at each center, gradually introduce the data of subsequent time batches for cyclic local validation and model iterative fine-tuning, gradually expanding the scale and diversity of the training data. Consider the dataset of each subsequent batch at each center as an incremental task stream corresponding to the pathological data distribution in a specific time period. Based on the feature similarity of the old and new data distributions, automatically enhance the retention intensity of historical parameters. Then, use the dynamic regularization coefficient to adjust the retention intensity of historical parameters in the federated continuous learning framework , where represents the baseline value of the dynamic regularization coefficient. (T - t) represents the difference between the current batch year and the historical year, and a is the decay coefficient, ensuring that the forgetting rate of the model for early data decreases year by year when the model is updated, in order to balance old and new knowledge and avoid catastrophic forgetting.

[0063] 5. Loss calculation module:

[0064] During the entire training process, the loss calculation module comprehensively calculates the loss from the data obtained by the federated aggregation module and the time-aware module of cyclic local validation . Before federated aggregation, the local model optimizes local parameters according to the fuzzy ratio labels. Based on the image patch classification task, calculate the cross-entropy for all training image patches with assigned class labels

[0065]

[0066] where is the magnification factor of the loss function, is the set of all training image patches with assigned class labels represents the probability that the j-th image patch is classified into the k-th class, is the training weight of the k-th class of training image patches. For pathological images, training and loss calculation are performed at magnification factors of 5, 10, and 20 respectively. In addition, through random upsampling or random downsampling, the training image patches in a training batch are all from the same pathological section. If the tissue ratio of the k-th class is lower than the ratio threshold. Then the ratio fitting loss function term for the k-th class label is defined as:

[0067]

[0068] where represents the ratio threshold of the k-th class label on the WSI, Denotes the average probability that the j-th slice is classified into k categories under different method multiples.

[0069] Then the overall proportion label loss is:

[0070]

[0071] Based on the cross-entropy loss and the proportion label loss, when performing cyclic local validation, the regularization term in the time-aware module is added to the loss calculation to retain the important nodes in the model and avoid the degradation of the recognition ability of the historical pathological data distribution after the model is updated:

[0072]

[0073] Where Denotes the trainable parameters of the optimal model in the historical training stage, which are used as anchor points to constrain the update direction of the current model's trainable parameter set To avoid the degradation of the recognition ability of the historical data distribution, Is a dynamic regularization coefficient for adjusting the weight of the old and new data distributions, and its value decays with the number of training rounds to gradually adapt to the evolving characteristics of the new data. Is the quantization parameter of the Fisher Information Matrix (FIM) The importance weight of the parameters in the historical model. The higher the importance of the parameters, the greater the retention strength in the regularization constraint, thus preferentially protecting the key pathological features.

[0074] Therefore, the global loss function is designed as:

[0075]

[0076] Instance data

[0077] The dataset used in this method is divided into two parts. The data in the first part comes from the First Affiliated Hospital of Zhejiang University School of Medicine, including 4,568 HE slides of 931 patients with primary liver cancer, with an average of 4.91 slides per person, and is approved for use by the ethics committee (No. 2018 - 115). The data in the second part comes from the hepatocellular carcinoma project of The Cancer Genome Atlas (TCGA - LIHC), including 354 HE slides of 337 patients with hepatocellular carcinoma, such as diagnostic histopathological slides, corresponding pathology reports, and clinical data including follow - up. The dataset from Zhejiang University First Hospital is regarded as one center (FAH - ZJUMS), and TCGA - LIHC is divided into two centers, the United States (LIHC - US) and Canada (LIHC - CA) according to the collection source. At the same time, FAH - ZJUMS is divided into three time batches of 2014, 2015, and 2016 according to the collection time, including 1,268, 1,523, and 1,777 slides respectively. The experimental results are as follows:

[0078] Table 1 Performance testing of data within and between centers

[0079] Table 2 Training and testing of multiple time batches of FAH - ZJUMS

[0080] On the other hand, the present invention also discloses a multi - center hepatocellular carcinoma tumor pathological micro - necrosis recognition method based on federated continuous learning. The implementation steps of this method specifically refer to the implementation process of each module of the above - mentioned multi - center hepatocellular carcinoma tumor pathological micro - necrosis recognition system based on federated continuous learning, and specifically include the following steps:

[0081] (1) Obtain the time - series image data of full - size hepatocellular carcinoma tumor pathological sections of each center;

[0082] (2) Perform foreground segmentation on the image data, remove the non - tumor background area, and generate a standardized image block group for the classification task;

[0083] (3) After each round of training, upload the local model weights of each center to the server. The server calculates the global weights using a weighted average strategy based on the data sample size according to the data scale of each center, and distributes them to all participating centers for model iteration;

[0084] (4) Design a time - series sensitive regularization constraint to dynamically adjust the weight update intensity of local training in each center;

[0085] (5) Comprehensively calculate the cross - entropy loss and proportion label fitting loss of the image block classification task, as well as the regularization term loss during the model iteration process.

[0086] Corresponding to the embodiments of the above-mentioned method for identifying hepatocellular carcinoma pathological micro-necrosis based on federated continuous learning, the present invention also provides an embodiment of a device for identifying hepatocellular carcinoma pathological micro-necrosis based on federated continuous learning.

[0087] See Figure 3 , an embodiment of a device for identifying hepatocellular carcinoma pathological micro-necrosis based on federated continuous learning provided by the embodiments of the present invention includes a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it is used to implement the method for identifying hepatocellular carcinoma pathological micro-necrosis based on federated continuous learning in the above embodiments.

[0088] The embodiment of the device for identifying hepatocellular carcinoma pathological micro-necrosis based on federated continuous learning provided by the present invention can be applied to any device with data processing capabilities. The device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented by software, or by hardware, or by a combination of software and hardware. Taking software implementation as an example, as a logically meaningful device, it is formed by the processor of any device with data processing capabilities reading the corresponding computer program instructions in the non-volatile memory into the memory for operation. From the hardware level, as Figure 3 shown, it is a hardware structure diagram of any device with data processing capabilities where the device for identifying hepatocellular carcinoma pathological micro-necrosis based on federated continuous learning provided by the present invention is located. In addition to Figure 3 the processor, memory, network interface, and non-volatile memory shown, the device with data processing capabilities where the device in the embodiment is located usually also includes other hardware according to the actual functions of the device with data processing capabilities, which will not be elaborated here.

[0089] The specific implementation process of the functions and roles of each unit in the above device can be specifically referred to the implementation process of the corresponding steps in the above method, which will not be elaborated here.

[0090] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiments described above are only illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they can be located in one place, or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present invention solution. Those of ordinary skill in the art can understand and implement it without creative work.

[0091] An embodiment of the present invention further provides a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements a method for identifying pathological micro-necrosis of hepatocellular carcinoma based on federated continuous learning in the above embodiment.

[0092] The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium may also be an external storage device of any device with data processing capabilities, such as a plug-in hard disk, a Smart Media Card (SMC), an SD card, a Flash Card, etc. equipped on the device. Further, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and may also be used to temporarily store data that has been output or will be output.

[0093] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the method for identifying pathological micro-necrosis of hepatocellular carcinoma based on federated continuous learning.

[0094] The above embodiments are used to explain the present invention, rather than limit the present invention. Any modification and change made to the present invention within the spirit and scope of the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A hepatocellular carcinoma pathological micro-necrosis recognition system based on federated continuous learning, characterized in that, The system includes: a dataset construction module for obtaining time-series image data of full-size hepatocellular carcinoma tumor pathological sections from multiple centers; a data preprocessing module for performing foreground segmentation on the image data, removing non-tumor background regions, and generating a standardized image block group for the classification task; a federated aggregation module for uploading the local model weights of each center to the server after each round of training. The server calculates the global weights using a weighted average strategy based on the data sample size according to the data scale of each center, and distributes them to all participating centers for model iteration; a time awareness module for designing time-series sensitive regularization constraints and dynamically adjusting the weight update intensity of local training at each center; a loss calculation module for comprehensively calculating the cross-entropy loss and proportional label fitting loss of the image block classification task, as well as the regularization term loss during the model iteration process.

2. The hepatocellular carcinoma pathological micro - necrosis recognition system based on federated continuous learning according to claim 1, wherein The image data is a full-size digital scanned image of a tumor tissue pathological section obtained by hematoxylin-eosin staining, and a proportional label vector is constructed for the pathological image.

3. The hepatocellular carcinoma pathological micro-necrosis recognition system based on federated continuous learning according to claim 1, characterized in that, The data preprocessing module calculates the thumbnail size of the pathological image at different magnification factors according to the set target image block overlap rate, and uses the difference distribution characteristics of the RGB channels and grayscale values of the thumbnail to segment the foreground of the tumor tissue, remove the tumor tissue background, perform pixel-level segmentation on the thumbnail, with each pixel point corresponding to an image block. For pixel points, they are retained and excluded according to the segmentation threshold of the Otsu threshold segmentation method.

4. The hepatocellular carcinoma pathological micro-necrosis recognition system based on federated continuous learning according to claim 1, wherein For the first batch of datasets of each center, the federated aggregation module performs noise-robust training on the local server and optimizes the local model parameters through a fuzzy proportional label learning strategy.

5. The hepatocellular carcinoma pathological micro-necrosis recognition system based on federated continuous learning according to claim 1, wherein, Regarding the subsequent batch of datasets of each center as an incremental task flow corresponding to the pathological data distribution in a specific time period, based on the feature similarity of the old and new data distributions, the retention intensity of historical key parameters is automatically enhanced: where (T - t) represents the difference between the current batch year and the historical year, and a is the attenuation coefficient, ensuring that the forgetting rate of the model for early data decreases year by year during model update.

6. The hepatocellular carcinoma pathological micro-necrosis recognition system based on federated continuous learning according to claim 1, wherein, The cross-entropy loss of the image block classification task includes the losses at different magnification factors of the image blocks, and through random upsampling or random downsampling, the training image blocks in a training batch are all from the same pathological section.

7. A hepatocellular carcinoma pathological micro-necrosis recognition system based on federated continuous learning according to claim 1, characterized in that, The specific regularization term loss is as follows: Among them, represents the optimal model parameters in the historical training stage, which are used as anchor points to constrain the update direction of the current parameters to avoid the degradation of the recognition ability of the historical data distribution, adjust the weights of the old and new data distributions, and its value decays with the number of training rounds to gradually adapt to the evolution characteristics of the new data; is the quantization parameter of the Fisher Information Matrix (FIM) The importance weight of the parameter in the historical model. The higher the importance of the parameter, the greater the retention strength in the regularization constraint, so as to preferentially protect the key pathological features.

8. A method for identifying pathological micro - necrosis of hepatocellular carcinoma based on federated continuous learning of the system according to any one of claims 1 - 7, characterized in that, The method includes the following steps: (1) Obtaining time-series image data of full-size hepatocellular carcinoma tumor pathological sections from each center; (2) Performing foreground segmentation on the image data, removing non-tumor background regions, and generating a standardized image block group for the classification task; (3) Uploading the local model weights of each center to the server after each round of training. The server calculates the global weights using a weighted average strategy based on the data sample size according to the data scale of each center, and distributes them to all participating centers for model iteration; (4) Designing time-series sensitive regularization constraints and dynamically adjusting the weight update intensity of local training at each center; (5) Comprehensively calculating the cross-entropy loss and proportional label fitting loss of the image block classification task, as well as the regularization term loss during the model iteration process.

9. A device for identifying pathological micro-necrosis of hepatocellular carcinoma based on federated continuous learning, comprising a memory and one or more processors, wherein executable code is stored in the memory, characterized in that, When the processor executes the executable code, it implements a method for identifying pathological micro-necrosis of hepatocellular carcinoma based on federated continuous learning as described in claim 8.

10. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements a method for identifying pathological micro-necrosis of hepatocellular carcinoma based on federated continuous learning as described in claim 8.

Images