Hepatocellular carcinoma pathological micronecrosis recognition system and method based on federated continuous learning

Abstract

The present invention discloses a system and method for identifying hepatocellular carcinoma pathological micronecrosis based on federated continuous learning. First, time-series image data of full-size hepatocellular carcinoma tumor pathological sections from multiple centers are acquired; foreground segmentation is performed on the image data, non-tumor background areas are removed, and a standardized image block group for classification tasks is generated; then, after each round of training, the local model weights of each center are uploaded to a server. The server calculates the global weights based on the data scale of each center using a weighted average strategy based on the data sample size, and distributes them to all participating centers for model iteration; and a time-sensitive regularization constraint is designed to dynamically adjust the weight update strength of each center's local training; finally, 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, are comprehensively calculated.

Description

Technical Field

[0001] The present invention relates to the technical field of pathological image recognition, and in particular to a system and method for identifying pathological micronecrosis in hepatocellular carcinoma based on federated continuous learning. Background Art

[0002] Histopathology image analysis is crucial for diagnosing many cancers. The process begins with a pathologist obtaining a representative tissue sample from the patient through a biopsy. These samples then undergo a series of preparation steps, including fixation, dehydration, paraffin removal, maceration, embedding, and sectioning, ultimately resulting in pathological slides. After staining and mounting, the slides are ready for detailed microscopic examination by pathologists. Hematoxylin and eosin (H&E) are the most commonly used stains in histopathology. These dyes interact specifically with specific structures or components within biological tissues, resulting in distinct colors for different structures within the tissue slide. This color variation helps reveal cell morphology, structure, and distribution, providing pathologists with a wealth of histological information that plays an indispensable role in clinical diagnosis. This detailed visual information enables pathologists to accurately diagnose the disease and plan appropriate treatment plans for patients. However, this examination is time-consuming and can be subjective. Previous studies in histopathology grading have demonstrated poor inter-pathologist agreement. Furthermore, pathologists may miss small areas of cancer.

[0003] In the context of the digital revolution, digital slide scanners convert pathology slides into digital pathology images (WSIs), facilitating subsequent applications in storage, teaching, and remote image reading. In recent years, digital pathology and computational pathology have rapidly developed with the help of scanning equipment, computing servers, and deep learning algorithms. WSI research for various diseases has yielded promising results in cancer diagnosis, subtype classification, and prognostic stratification. Computer-assisted diagnosis systems using deep learning algorithms can not only mitigate the shortcomings of pathologists but also improve diagnosis by screening out clearly benign pathology slides and providing quantitative characterization of suspicious areas.

[0004] Therefore, many computer-aided pathology diagnosis systems are based on deep learning models trained on various tasks and applied to clinical practice. However, the vast majority of computer-aided pathology diagnosis systems train their deep learning models based on data from a single medical center. This is because the distribution of patient characteristics across different medical centers in real-world clinical settings can be highly heterogeneous. Furthermore, different medical institutions use different preparation procedures, stains, and scanners when preparing digital pathology images, resulting in varying appearances of tissue sections that can affect the robustness and generalization capabilities of deep learning models. Furthermore, due to the large size of single WSI slices, a single medical institution's WSI dataset often measures hundreds of GB or even TB. Aggregating WSI data from multiple medical institutions for model training presents numerous practical challenges, such as storage and data transmission. The high level of privacy inherent in medical data also makes centralized training more difficult. In addition, since the distribution of the patient population will also change dynamically over time, in order to meet the actual needs of clinical practice for computer pathology-assisted diagnosis systems, it is necessary to conduct regular and continuous local evaluations of the model to further ensure the model's clinical reliability. Therefore, designing a federated continuous learning system that can effectively and accurately reflect clinical needs, and can use 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 pathology-assisted diagnosis systems, and can improve the generalization and robustness of computer pathology-assisted diagnosis systems to a certain extent.

[0005] To address this issue, many studies have attempted to utilize federated learning to improve the generalization capabilities of computer-aided pathology diagnosis systems across multiple centers. The technical solutions most closely aligned with those claimed in this patent are: ① Model optimization based on improved federated aggregation. This approach addresses data heterogeneity by improving the model aggregation mechanism of federated learning. Federated Proximal (FedProx) (LI T, SAHU AK, 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. This term acts as a regularizer, penalizing large deviations from the global model. Simply put, it constrains local updates, bringing them closer to the global model. Another approach, Stochastic Controlled Averaging in Federated Learning (SCAFFOLD) (Karimireddy SP, 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 on each client to estimate the global update direction and adjust local updates to reduce bias introduced by heterogeneous data. (2) Model alignment mechanisms for data silos.This technical solution constrains local models to address the differences in the distribution of multi-center medical data. SM Hosseini et al. proposed a proportionally fair federated learning (Prop-FFL) (HOSSEINI SM, 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.) that aims to overcome the bias in client considerations by introducing a second loss objective that rewards similar training losses for all clients while also taking into account the proportion of training samples for each client. Clients with fewer samples receive lower weights. S.Albarqouni et al. proposed the SiloBN method (SiloedFederated 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 computing BN statistics for any unseen dataset during inference.

[0006] Existing technologies similar to ① increase model generalization capabilities by improving federated learning aggregation methods, but they are difficult to adapt to the dynamic evolution characteristics of multi-center data and have the following problems:

[0007] 1. Static regularization is incompatible with dynamic data: FedProx uses fixed parameter constraints to force local models to align with the global model, but this cannot adapt to the temporal evolution of multi-center data distribution. When data distribution shifts over time, fixed constraints can easily lead model updates into a cycle of "underfitting and divergence."

[0008] 2. Incremental updates forget historical knowledge: Traditional federated aggregation focuses only on generalization across the central space and does not design a continuous learning mechanism in the time dimension. As a result, the model loses key pathological features learned previously due to catastrophic forgetting during incremental updates.

[0009] 3. Global aggregation ignores local time series 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 local data in each center, resulting in model performance degradation during long-term deployment.

[0010] Existing technologies similar to (2) use a static distribution alignment strategy to improve the model's cross-domain generalization ability for heterogeneous target populations. However, this over-reliance on the assumption of a fixed data distribution leads to the following problems:

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

[0012] 2. Insufficient task adaptability: The existing alignment mechanism only optimizes the consistency of parameter distribution between model layers and does not design a cross-center feature alignment strategy for specific diagnostic tasks. This results in the decoupling of the standardized feature space from the diagnostic task objectives and limited recognition accuracy. Summary of the Invention

[0013] The purpose of the present invention is to address the deficiencies of the existing technology and propose a multi-center hepatocellular carcinoma tumor pathology micronecrosis recognition system and method based on federated continuous learning.

[0014] The objectives of the present invention are achieved through the following technical solutions: In a first aspect, the present invention provides a multi-center hepatocellular carcinoma tumor pathology micronecrosis recognition system based on federated continuous learning, the system comprising:

[0015] A dataset construction module to acquire time-series image data of full-size HCC tumor pathology sections from each center;

[0016] The data preprocessing module is used to perform foreground segmentation on the image data, remove non-tumor background areas, and generate standardized image block groups for classification tasks;

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

[0018] The time-aware module is used to design timing-sensitive regularization constraints and dynamically adjust the weight update intensity of local training at each center;

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

[0020] Furthermore, 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.

[0021] Furthermore, the data preprocessing module calculates the thumbnail size of the pathology image at different magnifications 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 tumor tissue foreground, remove the tumor tissue background, and perform pixel-level segmentation on the thumbnail. Each pixel corresponds to an image block, and the pixels are retained or eliminated based on the segmentation threshold of the Otsu threshold segmentation method.

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

[0023] Furthermore, the datasets of subsequent batches of each center are regarded as incremental task flows. Corresponding to the pathology data distribution in a specific time period, the retention strength of historical parameters is automatically enhanced based on the feature similarity of the new and old data distributions:

[0024]

[0025] Where (Tt) represents the difference between the current batch year and the historical year, and a is the decay coefficient, which ensures that the forgetting rate of early data decreases year by year when the model is updated.

[0026] Furthermore, the cross-entropy loss of the image patch classification task includes the loss under different magnifications of the image patch, and through random upsampling or random downsampling, the training image patches in a training batch are all from the same pathological slide.

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

[0028]

[0029] in, Represents the optimal model parameters in the historical training phase, which serve as anchor points to constrain the current parameters The update direction of , to avoid the degradation of the ability to identify the distribution of historical data, Adjust the weight of the distribution of new and old data, whose value decays with each training round to gradually adapt to the evolution characteristics of the new data; is the Fisher Information Matrix (FIM) quantization parameter Importance weights in the historical model, parameters with higher importance are retained more strongly in the regularization constraints, thereby prioritizing the protection of key pathological features.

[0030] In a second aspect, the present invention further discloses a multi-center method for identifying micronecrosis in hepatocellular carcinoma tumor pathology based on federated continuous learning, the method comprising the following steps:

[0031] (1) Obtain time-series image data of full-size HCC tumor pathology sections 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, the local model weights of each center are uploaded to the server. The server calculates the global weights based on the data scale of each center using a weighted average strategy based on the data sample size and distributes them to all participating centers for model iteration;

[0034] (4) Designing timing-sensitive regularization constraints to dynamically adjust the weight update intensity of local training at each center;

[0035] (5) Comprehensively calculate the cross entropy loss and proportional label fitting loss of the image patch 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 micronecrosis of hepatocellular carcinoma based on federated continuous learning, comprising a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the executable code, it implements the method for identifying pathological micronecrosis of hepatocellular carcinoma based on federated continuous learning.

[0037] In a fourth aspect, the present invention further discloses a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the method for identifying pathological micronecrosis of hepatocellular carcinoma based on federated continuous learning is implemented.

[0038] Beneficial effects of the present invention:

[0039] (1) The present invention introduces dynamic regularization constraints instead of fixed constraints to ensure the strong generalization ability of the model in clinical settings when the data distribution shifts over time.

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

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

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

[0043] (5) This paper designs a multimodal attention fusion module, combines morphological prior knowledge with deep feature expression, and establishes a cross-center robust micronecrosis quantitative evaluation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

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

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

[0047] Figure 3 This is a structural diagram of a device for identifying pathological micronecrosis in hepatocellular carcinoma based on federated continuous learning. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described below with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] like Figure 1 As shown, the present invention provides a multi-center hepatocellular carcinoma tumor pathology micronecrosis recognition system based on federated continuous learning, which mainly includes a data set construction module, a data preprocessing module, a federated aggregation module, a time perception module and a loss calculation module.

[0050] The dataset module constructs a structured dataset containing full-size hepatocellular carcinoma tumor pathology sections based on multi-center dynamic time series data; the data preprocessing module adopts a multi-scale foreground segmentation strategy, combined with Otsu threshold segmentation and image block reconstruction technology, to remove non-tumor background areas and generate standardized image block groups for classification tasks, thereby improving data quality and model training efficiency and supporting standardized expression of heterogeneous data across centers; the federated aggregation module realizes the collaborative optimization of multi-center models through robust training and data volume weighted synchronous aggregation mechanism, alleviating the impact of heterogeneous data distribution offset on global generalization performance; the time perception module introduces dynamic regularization constraints to adaptively adjust the weights of new and old data, suppressing catastrophic forgetting and enhancing the model's continuous adaptability to dynamic data distribution; the loss calculation module integrates cross-entropy loss, proportional label fitting loss and time-aware regularization terms, and achieves a balanced improvement in micro-necrosis area recognition accuracy and historical knowledge retention 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 acquired data set is denoted as B, which contains n full-size digital scanned images (Whole Slide Image, WSI) of tumor tissue pathology sections obtained by Hematoxylin-Eosin (HE) staining, namely , the classification task targets a total of C categories, for the i-th pathological image , construct the scale label vector as: ;

[0053] in, is the ratio threshold of the kth class in the i-th sample image, and satisfies The ratio threshold refers to the lowest value of the ratio range when roughly estimated manually. centers, the total data set , among which, Datasets from the center Can be divided into sub-datasets according to the collection time , Represents the jth batch of data included. After training on the first batch of data sets, the model will be dynamically updated based on the new batch of data sets.

[0054] 2. Data preprocessing module:

[0055] A standardized preprocessing process for multi-center pathology data was used to perform standardized annotation of HE-stained pathology sections based on foreground segmentation and proportional label vectors, breaking through the heterogeneity limitations of multi-center data and reducing interference from non-tumor areas. The first step of preprocessing is foreground segmentation. The pathology image is taken at different magnifications. The thumbnail size of the pathology image is calculated according to the set target image block overlap rate, and the difference distribution characteristics of the thumbnail's RGB channels and grayscale values ​​are used to segment the tumor tissue foreground, remove the tumor tissue background, and perform pixel-level segmentation on the thumbnail. Each pixel corresponds to an image block, and the pixels are retained or eliminated based on the Otsu threshold segmentation method. For a single data set, the total data set after removing the background is finally obtained. ,Include sample packages, the i-th sample package contains example image block groups, denoted as , Represents the image block groups after foreground segmentation at different magnifications.

[0056] 3. Federal aggregation module:

[0057] Based on the image block classification problem, this paper adopts a federated synchronous parameter aggregation mechanism to achieve cross-center collaborative training. Based on the training architecture of federated continuous learning and the local loop verification strategy, it avoids the deviation of patient pathological characteristics due to time changes and meets the needs of real-world clinical scenarios. The specific process is as follows:

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

[0059] (2) Synchronous parameter aggregation: After each round of training, each center will aggregate the local model weights. , and upload to the central server simultaneously. The server is based on the data scale of each center. (i.e. the total number of image blocks at the i-th center), the global weight is calculated using a weighted average strategy based on the amount of data samples in, For the full sample size, 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 will It is distributed 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 perception module:

[0062] Through the federated continuous learning framework, we design time-sensitive regularization constraints, dynamically adjust the weight update intensity of local training in each center, introduce time-aware dynamic regularization coefficients in distributed incremental training, and combine data volume weighting strategies to achieve adaptive distribution of cross-center model weights. After the first batch of models in each center achieves aggregate training, we gradually introduce data from subsequent time batches for cyclic local verification and model iterative fine-tuning, gradually expand the scale and diversity of training data, and regard the data sets of subsequent batches of each center as incremental task flows, corresponding to the pathological data distribution in a specific time period. Based on the feature similarity of the distribution of new and old data, we automatically enhance the retention strength of historical parameters. Then, using the dynamic regularization coefficient Adjusting the retention strength of historical parameters in the federated continuous learning framework ,in, Represents the baseline value of the dynamic regularization coefficient. (Tt) represents the difference between the current batch year and the historical year. a is the decay coefficient, which ensures that the forgetting rate of earlier data decreases year by year when the model is updated, thereby balancing new and old knowledge and avoiding catastrophic forgetting.

[0063] 5. Loss calculation module:

[0064] During the entire training process, the loss calculation module calculates the loss by combining the data obtained by the federated aggregation module and the time-aware module of the cyclic local verification. Before federated aggregation, the local model learns and optimizes local parameters based on the fuzzy scale label. Based on the image patch classification task, the cross entropy is calculated for all training image patches that have been assigned category labels.

[0065]

[0066] in is the magnification The loss function under is the set of all training image patches that have been assigned class labels represents the probability that the j-th image block is classified into k categories, is the training weight for the k-th class training image patch. For pathology images, training and loss calculation are performed using magnifications of 5, 10, and 20, respectively. Furthermore, random upsampling or downsampling is performed to ensure that all training image patches in a training batch come from the same pathology slide. If the proportion of tissue in class k is below the proportion threshold, the proportion fitting loss function term for the k-th class label is defined as:

[0067]

[0068] in, Indicates the The ratio threshold of class labels on WSI, It represents the average probability that the jth slice is classified into k categories under different method multiples.

[0069] Then the overall proportional label loss is:

[0070]

[0071] On the basis of cross-entropy loss and proportional label loss, during cyclic local validation, the regularization term in the time-aware module is added to the loss calculation to retain important nodes in the model and avoid a decrease in the ability to recognize the distribution of historical pathological data after the model is updated:

[0072]

[0073] in Represents the optimal model trainable parameters in the historical training phase, which serves as an anchor to constrain the current model trainable parameter set The update direction of , to avoid the degradation of the ability to identify the distribution of historical data, It is a dynamic regularization coefficient used to adjust the weight of the distribution of new and old data. Its value decays with the training rounds to gradually adapt to the evolution characteristics of the new data. is the Fisher Information Matrix (FIM) quantization parameter Importance weights in the historical model, parameters with higher importance are retained more strongly in the regularization constraints, thereby prioritizing the protection of key pathological features.

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

[0075]

[0076] Instance Data

[0077] This method utilizes two datasets. The first dataset, collected from the First Affiliated Hospital of Zhejiang University School of Medicine, includes 4568 HE sections from 931 patients with primary liver cancer (4.91 sections per patient). The dataset was approved for use by the ethics committee (No. 2018-115). The second dataset, collected from the Cancer Genome Atlas Hepatocellular Carcinoma Project (TCGA-LIHC), includes diagnostic histopathology slides, corresponding pathology reports, and clinical data, including follow-up, from 337 patients with HCC. The ZJUMS dataset was considered a single center (FAH-ZJUMS), while the TCGA-LIHC dataset was divided into two centers based on collection source: the United States (LIHC-US) and Canada (LIHC-CA). Furthermore, the FAH-ZJUMS dataset was divided into three time-series batches based on acquisition time: 2014, 2015, and 2016, containing 1268, 1523, and 1777 sections, respectively. The experimental results are as follows:

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

[0079]

[0080] Table 2 FAH-ZJUMS multi-time batch training and testing

[0081]

[0082] On the other hand, the present invention also discloses a method for identifying micronecrosis in hepatocellular carcinoma tumor pathology based on federated continuous learning. The implementation steps of this method are specifically referred to the implementation process of each module of the multi-center hepatocellular carcinoma tumor pathology micronecrosis identification system based on federated continuous learning, and specifically include the following steps:

[0083] (1) Obtain time-series image data of full-size HCC tumor pathology sections from each center;

[0084] (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;

[0085] (3) After each round of training, the local model weights of each center are uploaded to the server. The server calculates the global weights based on the data scale of each center using a weighted average strategy based on the data sample size and distributes them to all participating centers for model iteration;

[0086] (4) Designing timing-sensitive regularization constraints to dynamically adjust the weight update intensity of local training at each center;

[0087] (5) Comprehensively calculate the cross entropy loss and proportional label fitting loss of the image patch classification task, as well as the regularization term loss during the model iteration process.

[0088] Corresponding to the aforementioned embodiment of a method for identifying pathological micronecrosis of hepatocellular carcinoma based on federated continuous learning, the present invention also provides an embodiment of a device for identifying pathological micronecrosis of hepatocellular carcinoma based on federated continuous learning.

[0089] See also Figure 3 An embodiment of the present invention provides a device for identifying pathological micronecrosis of hepatocellular carcinoma based on federated continuous learning, comprising a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the executable code, it is used to implement a method for identifying pathological micronecrosis of hepatocellular carcinoma based on federated continuous learning in the above embodiment.

[0090] The embodiment of the device for identifying hepatocellular carcinoma pathological micronecrosis based on federated continuous learning provided by the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located reading the corresponding computer program instructions in the non-volatile memory into the memory for execution. From the hardware level, if Figure 3 As shown in the figure, a hardware structure diagram of a device for identifying hepatocellular carcinoma pathological micronecrosis based on federated continuous learning provided by the present invention is provided in any device with data processing capability, except Figure 3 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0091] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0092] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present invention. A person of ordinary skill in the art can understand and implement the present invention without inventive work.

[0093] An embodiment of the present invention also provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the method for identifying pathological micronecrosis of hepatocellular carcinoma based on federated continuous learning in the above embodiment is implemented.

[0094] The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities described in any of the aforementioned 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. Furthermore, 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 is to be output.

[0095] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the method for identifying pathological micronecrosis of hepatocellular carcinoma based on federated continuous learning.

[0096] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modifications and changes made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A hepatocellular carcinoma pathological micronecrosis recognition system based on federated continuous learning, characterized by: The system includes: A dataset construction module for acquiring time-series image data of full-size hepatocellular carcinoma tumor pathology sections from multiple centers; The data preprocessing module is used to perform foreground segmentation on the image data, remove non-tumor background areas, and generate standardized image block groups for classification tasks; The federated aggregation module is used to upload the local model weights of each center to the server after each round of training. The server calculates the global weights based on the data scale of each center using a weighted average strategy based on the data sample size and distributes them to all participating centers for model iteration; The time-aware module is used to design timing-sensitive regularization constraints and dynamically adjust the weight update intensity of local training in each center. The data sets of subsequent batches of each center are regarded as incremental task flows, corresponding to the pathological data distribution in a specific time period. Based on the feature similarity of the new and old data distributions, the dynamic regularization coefficient is used. Automatically enhance the retention strength of historical key parameters: in, represents the baseline value of the dynamic regularization coefficient, (Tt) represents the difference between the current batch year and the historical year, and a is the decay coefficient, which ensures that the forgetting rate of early data decreases year by year when the model is updated; The loss calculation module is used to comprehensively calculate the cross entropy loss and proportional label fitting loss of the image patch classification task, as well as the regularization loss during the model iteration process; the regularization loss is specifically as follows: in, Represents the optimal model parameters in the historical training phase, which serve as anchor points to constrain the current parameters The update direction of , to avoid the degradation of the ability to identify the distribution of historical data, Adjust the weight of the distribution of new and old data, whose value decays with each training round to gradually adapt to the evolution characteristics of the new data; is the Fisher Information Matrix (FIM) quantization parameter Importance weights in the historical model, parameters with higher importance are retained more strongly in the regularization constraints, thereby prioritizing the protection of key pathological features.

2. A hepatocellular carcinoma pathological micronecrosis recognition system based on federated continuous learning according to claim 1, characterized in that: The image data are full-size digital scanned images of tumor tissue pathological sections obtained by hematoxylin-eosin staining, and a proportional label vector is constructed for the pathological image.

3. The hepatocellular carcinoma pathological micronecrosis recognition system based on federated continuous learning according to claim 1 is characterized in that: The data preprocessing module calculates the thumbnail size of the pathology image at different magnifications based on the set target image block overlap rate, and uses the difference distribution characteristics of the thumbnail's RGB channels and grayscale values ​​to segment the tumor tissue foreground, remove the tumor tissue background, and perform pixel-level segmentation on the thumbnail. Each pixel corresponds to an image block, and pixels are retained or eliminated based on the segmentation threshold of the Otsu threshold segmentation method.

4. The hepatocellular carcinoma pathological micronecrosis recognition system based on federated continuous learning according to claim 1 is characterized in that: The federated aggregation module performs noise robustness training on the local server for the first batch of data sets of each center and optimizes local model parameters through a fuzzy proportional label learning strategy.

5. The hepatocellular carcinoma pathological micronecrosis recognition system based on federated continuous learning according to claim 1 is characterized in that: The cross entropy loss of the image patch classification task includes the loss of the image patch at different magnifications, and through random upsampling or random downsampling, the training image patches in a training batch are all from the same pathological slide.

6. A method for identifying pathological micronecrosis in hepatocellular carcinoma based on federated continuous learning and the system according to any one of claims 1 to 5, characterized in that: The method comprises the following steps: (1) Obtain time-series image data of full-size HCC tumor pathology sections from each center; (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; (3) After each round of training, the local model weights of each center are uploaded to the server. The server calculates the global weights based on the data scale of each center using a weighted average strategy based on the data sample size and distributes them to all participating centers for model iteration; (4) Designing timing-sensitive regularization constraints to dynamically adjust the weight update intensity of local training at each center; (5) Comprehensively calculate the cross entropy loss and proportional label fitting loss of the image patch classification task, as well as the regularization term loss during the model iteration process.

7. A device for identifying micronecrosis in hepatocellular carcinoma based on federated continuous learning, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the processor executes the executable code, it implements the method for identifying pathological micronecrosis of hepatocellular carcinoma based on federated continuous learning as described in claim 6.

8. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by the processor, the method for identifying pathological micronecrosis of hepatocellular carcinoma based on federated continuous learning as described in claim 6 is implemented.

Images