Intelligent image data processing method and system based on reinforcement learning

Abstract

The application relates to the technical field of image data processing, and discloses an intelligent image data processing method and system based on reinforcement learning, which realizes adaptive and customized intelligent data processing of input liver pathological images by constructing an end-to-end reinforcement learning model integrating a context analysis network, a policy network and a reward calculation network.

Description

Technical Field

[0001] This application relates to the field of image data processing technology, and more specifically, to an intelligent image data processing method and system based on reinforcement learning. Background Technology

[0002] In the field of modern medical image analysis, especially in pathological diagnosis, precise processing of tissue slide images is a crucial step in ensuring diagnostic accuracy. Traditional pathological image preprocessing and enhancement, such as contrast adjustment, color correction, noise suppression, and feature enhancement of specific regions, often require experienced pathologists or physicians to manually set parameters based on the specific tissue type, staining method, and potential lesion characteristics. This process is not only time-consuming and labor-intensive, but also highly subjective due to differences in the experience of different operators, making standardization and high-throughput processing difficult. Especially when dealing with massive amounts of digital pathological slides, this manual-dependent approach is inefficient and prone to introducing errors.

[0003] While existing technologies have developed intelligent image data processing systems based on reinforcement learning that can abstract pathological image processing into a sequential decision problem and autonomously select and execute a series of image processing operations through interaction with the original pathological images to generate images most beneficial for subsequent diagnostic auxiliary models to perform accurate analysis, their performance significantly degrades when these systems are migrated from their initially optimized specific tasks (e.g., lung adenocarcinoma diagnosis) to new tasks with different tissue characteristics, diagnostic focuses, large fluctuations in image quality, the presence of specific physical artifacts, and where the diagnostic objective changes from recognition to precise grading (e.g., liver fibrosis grading assessment). Summary of the Invention

[0004] This invention provides an intelligent image data processing method and system based on reinforcement learning, aiming to solve the problem that the performance of existing intelligent image data processing systems based on reinforcement learning significantly declines when they are transferred from specific tasks to new tasks with different tissue characteristics, diagnostic focus, large fluctuations in image quality, the presence of specific physical artifacts, and the shift in diagnostic objectives from recognition to precise grading. It also addresses the technical challenges in liver pathology image processing, such as staining differences, physical artifact interference, and inaccurate grading assessment.

[0005] The technical solution of this application is as follows: In a first aspect, this application discloses an intelligent image data processing method based on reinforcement learning, comprising: The input liver pathology image is obtained and used as the current state of the reinforcement learning model. By using the context analysis network in the reinforcement learning model, the current state is analyzed to generate context description information. The iterative process is repeated, which includes: generating a customized image processing sequence based on contextual description information through the policy network in the reinforcement learning model; processing the input liver pathology image according to the image processing sequence to obtain an intermediate processed image; evaluating the intermediate processed image through the reward calculation network in the reinforcement learning model, calculating the corresponding reward signal, and updating the network parameters of the policy network according to the reward signal; stopping the iteration when the reward signal meets the preset termination condition or the number of iterations reaches the preset maximum number of iterations, and outputting the intermediate processed image generated in the current iteration as the final processed image; Among them, the reinforcement learning model is an end-to-end intelligent liver pathology image processing model that integrates a context analysis network, a policy network, and a reward calculation network. By iteratively optimizing the network parameters of the policy network, it achieves adaptive and customized intelligent data processing of the input liver pathology images, forming a complete reinforcement learning image data processing closed loop.

[0006] This application provides an end-to-end reinforcement learning image data processing closed loop through this technical solution, which can adaptively perform customized intelligent processing on liver pathology images. It effectively solves the problems of manual dependence, strong subjectivity, low efficiency in traditional methods, as well as the performance degradation of existing reinforcement learning systems in liver pathology image processing. It achieves standardization and high throughput in image processing and provides a high-quality image foundation for subsequent accurate diagnosis.

[0007] Furthermore, based on the above, this application also proposes that the steps of performing context analysis on the current state through the context analysis network in the reinforcement learning model to generate context description information include: performing microstructure analysis, staining bias analysis, and tissue background analysis on the input liver pathology image to obtain microstructure feature information, first sub-context description information, and second sub-context description information, respectively; fusing the microstructure feature information, first sub-context description information, and second sub-context description information to obtain multi-dimensional fused context information; and generating the final context description information based on the multi-dimensional fused context information.

[0008] Through this technical solution, the context analysis network can comprehensively understand the current state of an image from three dimensions: microstructure, staining bias, and tissue background, and generate multi-dimensional fused context information. This provides rich and accurate basis for the policy network to generate more precise and targeted image processing sequences, effectively avoiding misjudgment or insufficient processing that may be caused by single-dimensional analysis.

[0009] Furthermore, in some preferred embodiments, the step of performing microstructural analysis on the input liver pathology image to obtain microstructural feature information includes: By using a context analysis network, multi-scale directional texture features of the input liver pathology image are extracted; based on the multi-scale directional texture features, a microstructure connection network corresponding to the input liver pathology image is constructed. Analyze the connectivity parameters, orientation consistency parameters, and branch point density parameters of the microstructure network, and calculate the structural integrity score of the microstructure network based on the analysis results; Based on the structural integrity score, combined with the local context information of the region where the microstructure is located corresponding to the microstructure connection network, the type of microstructure is identified and the degree of image blurring of the microstructure is quantified. Based on the identification results of the microstructure and the degree of fuzziness of the microstructure, microstructure feature information is generated.

[0010] This technical solution extracts multi-scale texture features from microstructures, constructs connectivity networks, and assesses structural integrity. It also combines local contextual information for type identification and fuzziness quantification. This allows for more precise capture of microstructural features in liver pathological images, providing more accurate microscopic guidance for subsequent image processing. It is particularly helpful in identifying and enhancing subtle pathological changes.

[0011] Based on the above, this application further proposes a step for analyzing staining deviations in an input liver pathology image to obtain a first sub-context description information, including: dividing the input liver pathology image into multiple independent local regions using a context analysis network; extracting the color distribution features of each local region and calculating the corresponding cell density parameters and tissue regularity parameters; comparing the color distribution features with a preset physiological color variation range benchmark to identify whether there are staining deviations in the local region that exceed the physiological range; adjusting the weights of the identified staining deviations based on the cell density parameters and tissue regularity parameters of the local region to distinguish between physiological color variations and staining artifacts; and generating a first sub-context description information to characterize the staining state of the image based on the staining deviation identification results after weight adjustment.

[0012] This technical solution analyzes the color distribution, cell density, and tissue arrangement regularity of local areas, and combines physiological color variation benchmarks and weight corrections to effectively distinguish between physiological color variations and staining artifacts. This allows for more accurate identification and quantification of staining deviations in images, providing a precise basis for subsequent color correction and enhancement, and avoiding over-processing of normal physiological variations.

[0013] As an optional approach, the steps of performing tissue background analysis on the input liver pathology image to obtain second sub-context description information include: dividing the input liver pathology image into multiple independent local regions using a context analysis network; extracting tissue structure features for each local region, including cell density features and extracellular matrix distribution pattern features; identifying the tissue type corresponding to the local region based on the tissue structure features, including at least one of normal liver tissue, steatosis region, and inflammatory infiltration region; selecting a corresponding background feature evaluation strategy from a preset strategy library based on the identified tissue type, where the background feature evaluation strategy predefines tissue background feature evaluation parameters and judgment thresholds corresponding to different tissue types; quantifying and evaluating the tissue background features of the local region according to the selected background feature evaluation strategy to identify whether there are abnormal features in the tissue background of the local region; and generating second sub-context description information to characterize the tissue background state of the image based on the abnormal evaluation results of the tissue background features.

[0014] This technical solution extracts organizational structure features and identifies organizational types in local areas, and matches background feature evaluation strategies according to a preset strategy library. It can perform customized background feature evaluation for different organizational types, thereby more accurately identifying abnormal features in the background of image organization, providing more refined background information for subsequent image processing, and helping to avoid misprocessing of normal organizational background.

[0015] Based on the above, this application further proposes that the image processing sequence includes at least one of the following operations: adaptive histogram equalization, anisotropic diffusion filtering, morphological closing operation, contrast-limited adaptive histogram equalization, and Gaussian pyramid decomposition and reconstruction.

[0016] This technical solution introduces a variety of advanced image processing operations, and the policy network can flexibly select and combine these operations according to the context description information, thereby generating more diverse and adaptive image processing sequences, effectively addressing the complex image quality problems and pathological feature enhancement needs in liver pathology images. To enhance functionality, a reward computation network within a reinforcement learning model is used to evaluate intermediate-processed images. The steps for calculating the corresponding reward signal include: extracting multi-dimensional features from the intermediate-processed images using the reward computation network, including texture features, morphological features, and color distribution features; matching the extracted multi-dimensional features with a pre-defined pathological feature template library and a pre-defined artifact feature library to identify and classify key pathological features and artifact features within the intermediate-processed images; quantifying and calculating key pathological feature indices and artifact feature indices based on the identification and classification results; and calculating a comprehensive evaluation score based on the quantified key pathological feature indices, artifact feature indices, and pre-defined image evaluation rules, and outputting the comprehensive evaluation score as a reward signal.

[0017] Through this technical solution, the reward computation network can comprehensively evaluate the quality of intermediate processed images from multiple levels, such as multi-dimensional features, pathological features, and artifact feature recognition, and quantify them into a comprehensive evaluation score. This provides a more comprehensive and refined feedback signal for the optimization of the policy network, effectively guiding the model to generate images that better meet diagnostic needs.

[0018] Based on the above, this application further proposes that the steps for multi-dimensional feature extraction of intermediate processed images through a reward calculation network include: A reward computation network is used to perform multi-scale texture analysis on the intermediate-processed image, capturing texture details at different scales. Based on the acquired texture details, morphological structure analysis is performed on the intermediate-processed image to identify and quantify the width, branch point number, and connection pattern parameters of collagen fiber bundles. Combining the texture details and morphological structure analysis results, local color space transformation and data statistics are performed on the intermediate-processed image to obtain the color distribution parameters of each local region. The texture details, morphological structure parameters, and color distribution parameters are compared with a preset standard feature range to distinguish between feature distortion areas introduced by image processing and areas of true pathological change in the intermediate-processed image. This preset standard feature range corresponds to... The method includes: 1) Feature value ranges for different pathological stages and artifact types; 2) Quantifying the distortion level of feature distortion regions based on comparison results; 3) Local region feature difference analysis of intermediate processed images to identify similar pathological feature regions with similar texture, morphology, and color distribution characteristics; 4) Adaptively adjusting local contrast enhancement parameters and spatial filtering parameters for the identified similar pathological feature regions to improve the boundary clarity between different types of pathological features; 5) Feature separation processing of the parameter-adjusted similar pathological feature regions to highlight the morphological differences between early diffuse collagen fibers and mature fiber septa; 6) Quantitative statistical analysis of the image features after feature separation to generate highly distinguishable multi-dimensional features for subsequent comprehensive evaluation score calculation.

[0019] This technical solution utilizes multi-scale texture analysis, morphological structure analysis, local color space conversion, and feature distortion region quantification. Furthermore, it performs local region feature difference analysis and feature separation processing to generate highly distinguishable multi-dimensional features. This allows for a more refined evaluation of image processing effects, with significant advantages in distinguishing feature distortion from real pathological changes and improving the clarity of pathological feature boundaries. It provides a more reliable basis for the grading assessment of liver fibrosis.

[0020] Building upon the above, this application further proposes that the step of performing context analysis on the current state through a context analysis network in a reinforcement learning model to generate context description information further includes: Multi-scale directional texture features are extracted from input liver pathology images using a context analysis network. Based on these features, a microstructure connectivity network corresponding to the input liver pathology images is constructed. The connectivity components, directional consistency, and branch point density parameters of the microstructure connectivity network are analyzed to calculate the structural integrity score. Based on the structural integrity score and the local context information of the region where the microstructure connectivity network is located, the type of microstructure is identified, and the degree of ambiguity is quantified. The obtained microstructure identification results and the degree of ambiguity are subjected to multi-dimensional feature fusion processing to integrate the image microstructure feature information. Based on the fused microstructure feature information, complete context description information is generated.

[0021] This technical solution extracts multi-scale texture features from microstructures, constructs connection networks, assesses structural integrity, identifies types, and quantifies fuzziness, and performs multi-dimensional feature fusion. This allows for a more comprehensive and refined capture of the microstructural features of images, providing richer contextual information for the policy network to generate more accurate image processing sequences.

[0022] Secondly, this application also discloses a reinforcement learning-based intelligent image data processing system for executing the aforementioned reinforcement learning-based intelligent image data processing method. The system includes: The system comprises the following modules: a state acquisition module, which acquires the input liver pathology image and uses it as the current state of the reinforcement learning model; a context analysis module, which analyzes the current state using the context analysis network within the reinforcement learning model to generate context description information; a policy network module, which generates a customized image processing sequence based on the context description information using the policy network within the reinforcement learning model; an image processing module, which processes the input liver pathology image according to the image processing sequence to obtain an intermediate processed image; a reward calculation module, which evaluates the intermediate processed image and calculates the reward signal using the reward calculation network within the reinforcement learning model; a parameter update module, which updates the network parameters of the policy network based on the reward signal; and an iteration control module, which iteratively calls the policy network module, image processing module, reward calculation module, and parameter update module until the reward signal meets a preset termination condition or reaches a preset maximum number of iterations, at which point iteration stops and the current intermediate processed image is output as the final processed image.

[0023] This application provides a system capable of performing the above-mentioned methods through this technical solution. Through modular design, it achieves intelligent and adaptive processing of liver pathology images, effectively solving the limitations of traditional manual processing, and providing an efficient and accurate image preprocessing tool for liver pathology diagnosis. Beneficial effects

[0024] This application discloses an intelligent image data processing method and system based on reinforcement learning. By constructing an end-to-end reinforcement learning model integrating a context analysis network, a policy network, and a reward computation network, it achieves adaptive and customized intelligent data processing of input liver pathology images. The method first acquires the input liver pathology image as the current state. The context analysis network performs context analysis on this state, generating context description information to comprehensively understand key features such as the image's microstructure, staining deviations, and tissue background. Subsequently, the policy network generates a customized image processing sequence based on the context description information, processing the image to obtain an intermediate processed image. The reward computation network performs multi-dimensional evaluation of the intermediate processed image, calculates the reward signal, and updates the network parameters of the policy network based on the reward signal, forming a complete reinforcement learning closed loop. This iterative optimization process continues until the reward signal meets a preset termination condition or reaches the maximum number of iterations, ultimately outputting a high-quality pathology image.

[0025] Compared with the prior art, this application has the following advantages: Addressing the issue of performance degradation across tasks: To address the significant performance degradation of existing reinforcement learning systems when migrating from specific tasks (such as lung adenocarcinoma diagnosis) to new tasks such as liver fibrosis grading assessment, this application utilizes a context analysis network to perform multi-dimensional analysis of liver pathological images, including microstructure, staining bias, and tissue background. This generates detailed contextual description information, enabling the policy network to generate highly customized image processing sequences based on the unique tissue characteristics of the liver and diagnostic needs. This effectively avoids misjudgments and insufficient processing caused by tissue differences.

[0026] Overcoming staining differences and physical artifacts: The contextual analysis network in this application can analyze the color distribution, cell density, and tissue regularity of local areas. Combined with physiological color variation benchmarks and weight corrections, it effectively distinguishes between physiological color variations and staining artifacts, avoiding over-processing of normal physiological variations. Simultaneously, through meticulous evaluation of the tissue background, it can identify and process common physical artifacts in liver biopsy sections, preventing them from being confused with true pathological features, thereby improving image realism and diagnostic usability.

[0027] Improving the accuracy of liver fibrosis grading assessment: The reward computation network in this application extracts multi-dimensional features from intermediate processed images, including texture, morphological, and color distribution features, and matches them with pathological feature template libraries and artifact feature libraries to quantify key pathological and artifact feature indicators. More importantly, this reward mechanism can distinguish between feature distortion regions introduced by image processing and regions of true pathological changes, and adaptively adjusts and separates similar pathological feature regions, highlighting the morphological differences between early diffuse collagen fibers and mature fibrous septa. This generates highly discriminative multi-dimensional features, effectively guiding the agent to optimize specific morphological features crucial for liver fibrosis grading, significantly improving the accuracy of downstream grading models, and solving the problem that existing system reward mechanisms cannot effectively support accurate grading assessment.

[0028] Achieving standardization and high throughput in image processing: The end-to-end reinforcement learning model of this application realizes the automation and intelligence of image processing, gets rid of the dependence on human experience, and effectively solves the problems of time-consuming, labor-intensive, subjective, difficult to standardize and high-throughput processing in traditional methods. It is especially suitable for processing massive digital pathological slides, improving diagnostic efficiency and consistency. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating an intelligent image data processing method based on reinforcement learning provided in an embodiment of the present invention. Figure 2 This is a flowchart of a method for generating context description information provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent image data processing system based on reinforcement learning provided in an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Reference Figure 1 , Figure 1 This is a flowchart illustrating an intelligent image data processing method based on reinforcement learning provided in an embodiment of the present invention, including: S11, Obtain the input liver pathology image and use the input liver pathology image as the current state of the reinforcement learning model; S12, The current state is analyzed using the context analysis network in the reinforcement learning model to generate context description information; S13, repeat the iterative process, the iterative process includes: generating a customized image processing sequence based on the context description information through the policy network in the reinforcement learning model, processing the input liver pathology image according to the image processing sequence to obtain an intermediate processed image, evaluating the intermediate processed image through the reward calculation network in the reinforcement learning model, calculating the corresponding reward signal, and updating the network parameters of the policy network according to the reward signal; S14, until the reward signal meets the preset termination condition or the number of iterations reaches the preset maximum number of iterations, stop the iteration, and output the intermediate processed image generated in the current iteration as the final processed image; The reinforcement learning model is an end-to-end intelligent liver pathology image processing model that integrates the context analysis network, the policy network, and the reward calculation network. By iteratively optimizing the network parameters of the policy network, it achieves adaptive and customized intelligent data processing of the input liver pathology image, thus forming a complete reinforcement learning image data processing closed loop.

[0032] This application introduces a context analysis network to gain a deep understanding of the input image and generate a customized processing sequence based on it. This effectively overcomes the limitations of traditional methods in terms of processing efficiency, standardization, and accuracy. In particular, it can significantly improve processing performance and diagnostic assistance capabilities when dealing with complex and variable pathological image data.

[0033] The core of the reinforcement learning-based intelligent image data processing method proposed in this application lies in constructing an end-to-end reinforcement learning model that integrates a context analysis network, a policy network, and a reward computation network. The "reinforcement learning model" is a machine learning paradigm whose goal is to train an agent to maximize cumulative rewards through trial and error in a specific environment. In this application, the model is designed for intelligent processing of liver pathology images. The "context analysis network" is responsible for understanding the features and potential problems of the current image, generating "contextual description information," which is crucial for guiding subsequent image processing decisions. The "policy network," based on the contextual description information, generates a series of "customized image processing sequences," which are operation steps tailored to the characteristics of the current image. The "reward computation network" is used to evaluate the effect of image processing and generate a "reward signal," which serves as the basis for optimizing the policy network. The "iterative process" refers to the continuous execution of image processing, evaluation, and parameter updates by the model until a preset condition is reached. The "preset termination condition" could be that the reward signal reaches a certain threshold, indicating that the image processing effect is good enough; the "preset maximum number of iterations" is to prevent infinite iteration.

[0034] In practical implementation, the intelligent image data processing method of this application first acquires the input liver pathological image and uses it as the current state of the reinforcement learning model. For example, pathological slice images of liver tissue can be acquired through a digital scanner and converted into a digital format that the model can process, such as PNG, JPEG, or TIFF. These images may contain various pathological features, staining artifacts, or tissue background information.

[0035] Subsequently, the context analysis network in the reinforcement learning model performs context analysis on the current state, generating contextual description information. The context analysis network can be implemented using various deep learning architectures, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), or Transformers. This network receives the input liver pathology image and performs feature extraction and semantic understanding on it. For example, the context analysis network can analyze the overall color distribution, texture features, cell density, and tissue structure of the image to identify the presence of uneven staining, blurred areas, specific pathological structures (such as fibrosis or inflammation), or artifacts. The contextual description information can be a vector or tensor containing these analysis results, used to comprehensively characterize the features of the current image and processing requirements.

[0036] Next, the system repeats the iterative process. In each iteration, the policy network generates a customized image processing sequence based on the context description information. The policy network can also employ a deep learning architecture, such as a Transformer-based sequence generation model or an RNN-based policy network. This network outputs a series of image processing operations and their parameters based on the context description information. For example, if the context description information indicates that the image has insufficient local contrast and slight noise, the policy network might generate a sequence containing "local histogram equalization" and "Gaussian filtering." The operations in the image processing sequence can be any combination from a predefined operation library, such as brightness adjustment, contrast enhancement, color correction, sharpening, denoising, morphological operations, etc.

[0037] Then, the input liver pathology image is processed according to the generated image processing sequence to obtain an intermediate processed image. For example, if the sequence contains "adaptive histogram equalization" and "anisotropic diffusion filtering", the image will be processed by these two operations in sequence. These operations can be implemented by functions in image processing libraries (such as OpenCV and Pillow).

[0038] Next, the reward computation network evaluates the intermediate-processed image and calculates the corresponding reward signal. The reward computation network can be an independent deep learning model, taking the intermediate-processed image as input and outputting a quantified reward value. For example, the reward computation network can evaluate image sharpness, contrast, the prominence of pathological features, and the degree of artifact suppression. The design of the reward signal is crucial, as it directly influences the learning direction of the policy network. For instance, if the processed image is sharper, pathological features are more easily identifiable, and artifacts are fewer, the reward signal will be higher.

[0039] The network parameters of the policy network are updated based on the reward signal. This is typically achieved using reinforcement learning algorithms such as Q-learning, SARSA, Policy Gradient, or Actor-Critic. The reward signal guides the policy network to adjust its internal parameters so that it can generate image processing sequences with higher rewards in the future. For example, if a processing sequence receives a high reward, the policy network will strengthen its tendency to generate that sequence; conversely, if it receives a low reward, it will weaken its tendency to generate that sequence.

[0040] The iterative process continues until the reward signal meets a preset termination condition or the number of iterations reaches a preset maximum number of iterations. The preset termination condition can be set as the reward signal reaching a certain threshold multiple times consecutively, indicating that the image processing effect has stabilized and optimized. The preset maximum number of iterations is an upper limit to prevent the model from getting stuck in an infinite loop. When either condition is met, the iteration stops, and the intermediate processed image generated in the current iteration is output as the final processed image.

[0041] The reinforcement learning model presented in this application is an end-to-end intelligent processing model for liver pathology images, integrating a context analysis network, a policy network, and a reward computation network. By iteratively optimizing the network parameters of the policy network, adaptive and customized intelligent data processing of input liver pathology images is achieved, forming a complete reinforcement learning image data processing closed loop. This end-to-end integration approach enables the model to directly learn the optimal processing strategy from the original image without manual intervention, thereby significantly improving processing efficiency and effectiveness.

[0042] The core innovation of this application lies in its end-to-end reinforcement learning framework. This framework uses a context analysis network to gain a deep understanding of pathological images, and on this basis, a policy network generates customized image processing sequences. Finally, a reward computation network quantifies the processing effect and guides policy optimization. Compared with traditional image processing methods that rely on expert experience and manual parameter setting, this application can achieve automation, standardization, and personalization of image processing. Compared with existing reinforcement learning-based image processing systems, this application, by introducing a context analysis network, can more precisely understand the complex context of liver pathological images, including microstructure, staining bias, and tissue background, thereby generating more targeted and effective processing strategies. This deep context awareness capability enables the model to effectively avoid artifacts and blurred key features when processing liver slices from different sources and at different pathological stages, significantly improving the ability to identify and classify subtle features such as liver fibrosis.

[0043] Furthermore, the design of the reward computation network places greater emphasis on the identification and quantification of key pathological features and artifacts, ensuring that the processing results can better serve the needs of accurate diagnosis and grading, thus demonstrating significant progress in processing efficiency, accuracy, and robustness.

[0044] For details, please refer to Figure 2 , Figure 2 This is a flowchart of a method for generating context description information provided in an embodiment of the present invention. S12 includes: S121, perform microstructure analysis, staining deviation analysis and tissue background analysis on the input liver pathology image to obtain microstructure feature information, first sub-context description information and second sub-context description information respectively; S122, the microstructure feature information, the first sub-context description information, and the second sub-context description information are fused to obtain multi-dimensional fused context information, and the final context description information is generated based on the multi-dimensional fused context information.

[0045] Microstructural analysis refers to the in-depth identification and quantification of subtle structural features in liver pathological images, such as cell morphology, tissue arrangement patterns, and degree of fibrosis. Through the analysis of these microstructures, information reflecting the core characteristics of pathological changes can be obtained, such as cell nucleus size, nuclear-cytoplasmic ratio, and the width and distribution of collagen fiber bundles.

[0046] Staining bias analysis assesses the uniformity and intensity of stain distribution in pathological images, as well as the presence of non-physiological color variations. During the acquisition and preparation of pathological images, the staining process may introduce biases such as uneven staining, overstaining, or understaining, which can affect the accurate interpretation of the images. Staining bias analysis generates first-subcontext descriptive information to characterize the staining quality and potential artifacts in the image.

[0047] Tissue background analysis refers to the assessment of the overall tissue environment of an input liver pathology image to identify the presence of background lesions such as fatty degeneration, inflammatory infiltration, necrosis, and fibrosis. This background information is crucial for understanding the overall context of the pathology image, as it can influence the identification of lesion areas and the selection of image processing strategies. Tissue background analysis can generate second sub-contextual description information to describe the macroscopic histopathological background of the image.

[0048] Furthermore, microstructural feature information, first sub-context description information, and second sub-context description information are fused to form multi-dimensional fused contextual information. This fusion aims to integrate pathological image features from different levels, thereby providing a comprehensive, detailed, and accurate image contextual description. Based on this multi-dimensional fused contextual information, the final contextual description information is generated, which will serve as the basis for the policy network in the reinforcement learning model to generate customized image processing sequences.

[0049] This application's approach refines contextual analysis into microstructural analysis, staining bias analysis, and tissue background analysis, enabling a comprehensive understanding of the complexity of input liver pathology images from multiple dimensions. Microstructural analysis focuses on the core diagnostic features of the pathological image, ensuring accurate capture of lesion details; staining bias analysis identifies and quantifies non-pathological factors that may affect image quality and interpretation accuracy; and tissue background analysis provides macroscopic pathological environment information of the image. By comprehensively considering and integrating these different dimensions of information, a richer and more accurate contextual description can be generated. This multi-dimensional contextual description can more comprehensively reflect the true state and potential problems of the image, thus providing a solid foundation for the subsequent strategy network to generate more targeted and effective image processing sequences.

[0050] Through the aforementioned technical solutions, the contextual analysis network can achieve a more comprehensive, detailed, and accurate understanding of liver pathological images. Specifically, by distinguishing microstructural features, staining biases, and tissue background, it can avoid misjudgments or information omissions that may result from single-dimensional analysis. This multi-dimensional contextual description information enables the reinforcement learning model to more accurately identify various problems in the image (such as blurring, artifacts, and specific pathological features), and generate highly customized image processing sequences accordingly. This significantly improves the adaptability and effectiveness of image processing, ultimately resulting in higher-quality, more easily diagnosed pathological images.

[0051] In some embodiments of this application, microstructural analysis of the input liver pathology image is a key step in context analysis. Specifically, the steps of performing microstructural analysis on the input liver pathology image to obtain microstructural feature information include: Multi-scale directional texture features are extracted from input liver pathology images using a context analysis network. Based on multi-scale directional texture features, a microstructure connection network corresponding to the input liver pathology image is constructed. Analyze the connectivity parameters, orientation consistency parameters, and branch point density parameters of the microstructure network, and calculate the structural integrity score of the microstructure network based on the analysis results; Based on the structural integrity score, combined with the local context information of the region where the microstructure is located corresponding to the microstructure connection network, the type of microstructure is identified and the degree of image blurring of the microstructure is quantified. Based on the identification results of the microstructure and the degree of fuzziness of the microstructure, microstructure feature information is generated.

[0052] Specifically, when extracting multi-scale directional texture features from input liver pathology images, various image processing techniques can be employed. For example, by applying a series of Gabor filter banks or wavelet transforms at different scales and directions, texture information at different frequencies and directions can be captured from the image. These texture features can effectively characterize the periodicity, directionality, and coarseness of microstructures such as cell arrangement and fibrosis degree in liver tissue.

[0053] Furthermore, based on the extracted multi-scale directional texture features, a microstructural connectivity network corresponding to the input liver pathology image can be constructed. This network can be understood as a graph structure, where nodes represent key microstructural units in the image, such as cell nuclei, cytoplasmic regions, and collagen fiber bundles, while edges represent the spatial proximity or functional connectivity between these units. The construction process may include clustering or thresholding the texture features to identify potential microstructural units and establishing connections based on their distance or feature similarity.

[0054] The analysis of the microstructure connection network involves several key parameters. The connectivity component parameter assesses the number and size of independently connected subgraphs within the network, reflecting the overall connectivity and dispersion of the microstructure. The orientation consistency parameter quantifies the regularity or disorder of the microstructure arrangement by analyzing the directional distribution of edges in the network; for example, it can be characterized by calculating the entropy or variance of fiber orientations within local regions. The branch point density parameter is the ratio of the number of points connecting three or more nodes to the total number of nodes in the network; it reflects the complexity and branching degree of the microstructure. For example, in liver fibrosis, excessive branching of collagen fibers may lead to an increase in branch point density. Based on the analysis results of these parameters, a structural integrity score of the microstructure connection network can be calculated, which comprehensively reflects the health status or pathological changes of the microstructure.

[0055] As a preferred implementation, when identifying the type of microstructures, machine learning classifiers, such as support vector machines or deep neural networks, can be used, with structural integrity scores as one of the input features. Simultaneously, combining local contextual information of the region where the microstructure is located, such as cell density, inflammatory infiltration, or degree of fatty degeneration in the surrounding area, with the microstructure's connectivity network, can improve the accuracy of identification. The quantification of image blur can be achieved by analyzing the gradient intensity, contrast, or high-frequency components of the Fourier transform at the edges of the microstructures to assess the blurring effect introduced during image acquisition or processing.

[0056] Therefore, based on the identification results of microstructures, such as normal liver lobules, fibrous septa, and fatty degeneration areas, as well as the quantified degree of ambiguity of the microstructures, microstructural feature information can be generated for subsequent contextual analysis. This information can be a multi-dimensional vector containing microstructure type, structural integrity score, degree of ambiguity, and other relevant quantitative indicators, providing reinforcement learning models with a comprehensive and accurate description of the microscopic pathological state of the image.

[0057] This application's solution extracts multi-scale directional texture features from input liver pathology images, comprehensively capturing structural information at different levels of detail. Based on these texture features, a microstructural connectivity network is constructed, abstracting the image's microstructure into a quantifiable network topology. This allows for in-depth evaluation of the tissue morphology and pathological changes of the microstructure by analyzing connected component parameters, directional consistency parameters, and branch point density parameters. The calculation of structural integrity scores provides a unified measure for quantifying the health status of microstructures. Furthermore, local contextual information is incorporated to identify microstructure types and quantify image blur, ensuring the accuracy of microstructure identification and consideration of image quality. The resulting microstructural feature information provides a refined, multi-dimensional micropathological basis for reinforcement learning models, enabling them to more accurately understand the current image state and generate more targeted image processing sequences.

[0058] Specifically, in the step of performing context analysis on the current state through the context analysis network in the reinforcement learning model to generate context description information, the step of performing staining bias analysis on the input liver pathology image to obtain the first sub-context description information may include the following specific operations: The input liver pathology image is divided into multiple independent local regions by region segmentation. For each of the local regions, the color distribution features of the local region are extracted, and the cell density parameters and tissue arrangement regularity parameters corresponding to the local region are calculated. The color distribution characteristics are compared with a preset physiological color variation range benchmark to identify whether there is a staining deviation in the local area that exceeds the physiological range. Based on the cell density parameters and tissue arrangement regularity parameters of the local area, the identified staining deviations are weighted and corrected to distinguish between physiological color variations and staining artifacts. Based on the weighted correction of the staining deviation identification results, a first sub-context description information is generated to characterize the staining state of the image.

[0059] The process involves segmenting the input liver pathology image into multiple independent local regions, which can be achieved using image segmentation algorithms. For example, deep learning-based semantic segmentation networks (such as U-Net and Mask R-CNN) or traditional image processing methods (such as watershed algorithms and region growing methods) can be used to perform pixel-level classification or region clustering of the image, thereby dividing the image into independent regions with different histological significance.

[0060] Furthermore, for each local region, its color distribution characteristics can be extracted and quantitatively described using methods such as color histograms, color moments (e.g., mean, variance, skewness, kurtosis), or color co-occurrence matrices. Simultaneously, the cell density parameters and tissue regularity parameters corresponding to the local region can be calculated using morphological image processing techniques. For example, cell density parameters can be obtained by statistically analyzing the ratio of the number of cell nuclei to the area of ​​the local region; tissue regularity parameters can be calculated by analyzing the geometry, orientation, and spatial relationships of cell nuclei or tissue structures. For instance, texture analysis methods such as gray-level co-occurrence matrices (GLCM) or local binary pattern (LBP) can be used to characterize tissue regularity.

[0061] Specifically, the color distribution characteristics are compared with a preset physiological color variation range benchmark to identify whether there are staining abnormalities in local areas that exceed the normal physiological range. The physiological color variation range benchmark is obtained based on statistical analysis of a large number of normal liver pathological image samples, covering the normal color fluctuation range caused by factors such as different staining batches and different slice thicknesses. The comparison process can use statistical tests (such as Z-score, T-test) or machine learning classifiers (such as support vector machines, neural networks) to determine whether the color characteristics of the current local area fall within the preset normal range.

[0062] Based on this, the identified staining deviations are weighted and corrected according to the cell density and tissue regularity parameters of the local area. The aim is to distinguish between physiological color variations and staining artifacts. For example, certain physiological changes (such as fatty degeneration or inflammatory cell infiltration) may cause changes in the color distribution of a local area, but its cell density and tissue regularity may still be within the normal or acceptable range. Staining artifacts (such as uneven staining, overstaining, or understaining) may be accompanied by significant abnormalities in cell density or tissue regularity. By introducing these structural parameters for weighting and correction, the risk of physiological variations being misjudged as staining artifacts can be effectively reduced, improving the accuracy of staining deviation identification.

[0063] Finally, based on the weighted and corrected color deviation identification results, a first sub-contextual description is generated to characterize the color deviation state of the image. This information can be a multi-dimensional vector containing the degree and type of color deviation (e.g., reddish, bluish, or abnormal brightness) for each local region, as well as the corrected confidence level. This information will serve as part of the contextual description to guide subsequent image processing strategies.

[0064] The proposed solution performs fine-grained region segmentation on the input liver pathology image and extracts multi-dimensional features for each local region, including color distribution, cell density, and tissue arrangement regularity, thereby comprehensively capturing the staining state of the image. By comparing the color features with a preset physiological color variation range benchmark, potential staining abnormalities can be preliminarily identified.

[0065] In some embodiments of this application described above, context analysis is performed on the input liver pathology image to generate context description information, which includes performing tissue background analysis on the input liver pathology image to obtain second sub-context description information. Specifically, the step of performing tissue background analysis on the input liver pathology image to obtain second sub-context description information includes: The context analysis network is used to segment and divide the input liver pathology image into multiple independent local regions. For each local region, the tissue structure features of the local region are extracted, including cell density features and extracellular matrix distribution pattern features. Based on the characteristics of the tissue structure, the tissue type corresponding to the local area is identified. The tissue type includes at least one of normal liver tissue, fatty degeneration area, and inflammatory infiltration area. Based on the identified organization type, the corresponding background feature evaluation strategy is selected from the preset strategy library. The background feature evaluation strategy has predefined organizational background feature evaluation parameters and judgment thresholds for different organization types. Based on the selected background feature evaluation strategy, the organizational background features of the local area are quantitatively evaluated to identify whether there are abnormal features in the organizational background of the local area; based on the abnormal evaluation results of the organizational background features, a second sub-contextual description information is generated to characterize the organizational background state of the image.

[0066] The process of segmenting input liver pathology images aims to decompose the complex overall image into more easily analyzable independent local regions. These local regions can be automatically segmented based on image texture, color, morphological features, etc., for example, through deep learning models or traditional image processing algorithms (such as superpixel segmentation, thresholding, etc.). For each local region, extracting tissue structure features is crucial. Cell density features reflect the density of cell numbers, for example, by calculating the number of cell nuclei or the proportion of cytoplasm per unit area. Extracellular matrix distribution pattern features focus on the morphology, density, and arrangement of the intercellular matrix, for example, through texture analysis (such as gray-level co-occurrence matrix, local binary pattern) or morphological operations (such as skeletonization, connected component analysis).

[0067] Based on the extracted tissue structure features, the tissue type of a local area can be identified. For example, high cell density and regularly arranged hepatocyte cords may indicate normal liver tissue; areas with numerous vacuolar structures and cell nuclei pushed to the edges may be identified as steatotic areas; while areas with aggregation of immune cells such as lymphocytes and plasma cells may be identified as inflammatory infiltration areas. This identification can be achieved through a trained classifier (such as a support vector machine or neural network) or a rule-based expert system. Once the tissue type is identified, the system will match and select the corresponding background feature evaluation strategy from a pre-defined strategy library. This pre-defined strategy library stores predefined tissue background feature evaluation parameters and judgment thresholds for different tissue types (such as normal liver tissue, steatotic areas, and inflammatory infiltration areas). For example, for normal liver tissue, the evaluation parameters may focus on the regularity of cell arrangement and the consistency of cell nucleus size; for steatotic areas, the size, number, and uniformity of fat vacuoles may be considered; and for inflammatory infiltration areas, the density, extent, and type of inflammatory cells may be considered.

[0068] The selected background feature evaluation strategy was then used to quantitatively evaluate the tissue background features of local regions. This included comparing the extracted tissue structure features with the evaluation parameters defined in the strategy and identifying whether there were abnormal features in the tissue background of local regions based on a judgment threshold. Abnormal features may include disordered cell arrangement, nuclear atypia, abnormal fibrosis, necrotic areas, etc. Finally, based on the anomaly evaluation results of the tissue background features, a second subcontextual description information was generated to characterize the tissue background state of the image. This information can be a vector, a text description, or structured data, detailing the tissue type of each local region in the image, the presence of anomalies, and the degree and nature of the anomalies, providing important contextual basis for subsequent image processing decisions.

[0069] The solution presented in this application, through refined tissue background analysis of input liver pathology images, enables a comprehensive and accurate understanding of the macroscopic pathological background of the images. Through the aforementioned technical solution, this application achieves refined and intelligent analysis of the tissue background of liver pathology images.

[0070] In some embodiments described above in this application, customized image processing sequences are generated based on contextual description information. However, in actual liver pathology image processing, the specific composition of the image processing sequence has a significant impact on the final processing effect. If the sequence is not properly selected, it may fail to effectively solve problems such as noise, artifacts, or insufficient contrast in the image, thereby affecting the accuracy of subsequent pathological diagnosis. Therefore, this application further proposes that in the aforementioned reinforcement learning-based intelligent image data processing method, the image processing sequence includes at least one of the following operations: adaptive histogram equalization, anisotropic diffusion filtering, morphological closing operation, contrast-limited adaptive histogram equalization, and Gaussian pyramid decomposition and reconstruction.

[0071] Specifically, adaptive histogram equalization involves dividing an image into multiple small regions and performing histogram equalization independently on each region to enhance local contrast. Its aim is to improve the overall visual quality and information recognizability of the image without sacrificing image details. Anisotropic diffusion filtering can be understood as a nonlinear filtering technique that smooths flat areas while preserving or enhancing edge information in edge regions. Specifically, it effectively removes image noise while maintaining the clarity of pathological tissue structures, aiming to avoid blurring important pathological features during noise reduction. Morphological closing operations involve performing dilation followed by erosion on the image to fill small holes, connect broken edges, or smooth image contours. Its purpose is to eliminate small noise points or connect broken tissue structures, making the morphological features of the pathological image more complete and continuous.

[0072] Contrast-limited adaptive histogram equalization is an improved form of adaptive histogram equalization. By limiting the magnitude of contrast enhancement, it avoids excessive enhancement that could lead to noise amplification or artifacts. Its aim is to effectively suppress noise and artifact generation while enhancing image contrast. Gaussian pyramid decomposition and reconstruction involves decomposing an image into image layers of different scales, processing at each scale, and then reconstructing the processed image layers to achieve multi-scale image analysis and processing. Its purpose is to refine image features at different resolutions, thereby better capturing and enhancing multi-scale information in pathological images.

[0073] The proposed solution incorporates operations such as adaptive histogram equalization, anisotropic diffusion filtering, morphological closing operation, contrast-limited adaptive histogram equalization, and Gaussian pyramid decomposition and reconstruction into the image processing sequence. This enables the reinforcement learning model to flexibly select and combine these operations based on contextual description information, thereby specifically addressing issues such as insufficient contrast, noise interference, incomplete structure, or blurred details in the input liver pathology images.

[0074] In response, this application further proposes a step for evaluating intermediate processed images and calculating the corresponding reward signals using a reward calculation network in a reinforcement learning model, including: A reward calculation network is used to extract multi-dimensional features from intermediate processed images. These multi-dimensional features include texture features, morphological features, and color distribution features. The extracted multi-dimensional features are matched with a preset pathological feature template library and a preset artifact feature library to complete the identification and classification of key pathological features and artifact features in the intermediate processed image. Based on the identification and classification results of the key pathological features and the artifact features, the key pathological feature index and the artifact feature index are quantitatively calculated respectively. Based on the quantitative indicators of key pathological features and the quantitative indicators of artifact features, combined with the preset image evaluation rules, a comprehensive evaluation score is calculated, and the comprehensive evaluation score is output as the reward signal.

[0075] Specifically, multi-dimensional feature extraction refers to acquiring various types of information from intermediate processed images to comprehensively evaluate image quality and content. Among them, texture features can be understood as patterns of pixel intensity changes in an image, such as edges, corners, and lines. Its purpose is to capture the fine structure and details of the image, which is crucial for identifying microstructures such as cell arrangement and degree of fibrosis in liver pathology images.

[0076] Morphological features refer to attributes such as shape, size, and topological structure of objects in an image. Their purpose is to quantify the geometric characteristics of pathological tissue structures, such as the size, shape, and arrangement of cell nuclei, as well as the morphology of collagen fiber bundles. Color distribution features refer to the proportion and spatial distribution of different color components in an image. Their purpose is to assess staining uniformity, cell staining intensity, and color differences between different tissue regions, which is of great significance for identifying staining deviations and distinguishing different pathological areas.

[0077] In practical applications, a pre-defined pathological feature template library can contain typical pathological feature patterns of various known liver diseases, such as the texture, morphology, and color manifestations of different degrees of steatosis, inflammatory infiltration, fibrosis, and hepatocellular carcinoma. A pre-defined artifact feature library can contain common image processing artifacts, such as edge halos caused by over-sharpening, color distortion caused by contrast enhancement, and noise amplification. By performing feature matching between the extracted image features and these template libraries, it is possible to accurately identify whether the target pathological features and artifacts introduced by the processing exist in the intermediate processed image.

[0078] Key pathological feature indicators can be quantitative assessments of identified pathological features, such as the salience, completeness, and clarity of the features. Artifact feature indicators can be quantitative assessments of identified artifacts, such as their intensity, extent, and degree of obscuring of pathological features. The calculation of these indicators can be based on commonly used quantitative methods in image processing, such as variations of the Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR), or specially designed pathological image quality assessment indicators. Predefined image evaluation rules define how to integrate these quantitative indicators to calculate the final comprehensive evaluation score. For example, weighted summation or multi-objective optimization functions can be used to ensure that the score accurately reflects the diagnostic value and visual quality of the image.

[0079] The solution proposed in this application addresses the limitations of traditional single-indicator evaluation of reward signals by introducing multi-dimensional feature extraction, identification and quantification of pathological and artifact features, and reward signal calculation based on comprehensive evaluation rules.

[0080] This application further proposes the following steps for multi-dimensional feature extraction of intermediate processed images using a reward computation network: By using a reward computation network, multi-scale texture analysis is performed on the intermediate processed image to capture texture details at different scales. Based on the acquired texture detail information, morphological structure analysis is performed on the intermediate processed image to identify and quantify the width parameters, branch point number parameters, and connection mode parameters of collagen fiber bundles in the intermediate processed image. By combining texture detail information and morphological structure analysis results, local color space conversion and data statistics are performed on the intermediate processed image to obtain the color distribution parameters of each local region. The texture detail information, morphological structure parameters, and color distribution parameters are compared with the preset standard feature range to distinguish the feature distortion areas introduced by image processing and the real pathological change areas in the intermediate processed image. The preset standard feature range corresponds to the feature value range of different pathological stages and different artifact types. Based on the comparison results, the degree of distortion in the quantified feature distortion region is determined. The method further includes: performing local region feature difference analysis on the intermediate processed image to identify similar pathological feature regions with similar texture, morphology, and color distribution characteristics within the image; adaptively adjusting the local contrast enhancement parameters and spatial filtering parameters for the identified similar pathological feature regions to improve the boundary clarity between different types of pathological features; performing feature separation processing on the similar pathological feature regions after parameter adjustment to highlight the morphological differences between early diffuse collagen fibers and mature fiber intervals; and performing quantitative statistics on the image features after feature separation to generate highly distinguishable multi-dimensional features for subsequent comprehensive evaluation score calculation.

[0081] Specifically, multi-scale texture analysis aims to capture texture patterns in images from coarse to fine using filters or transforms of different scales (such as wavelet transform, Gabor filter banks, etc.), thereby comprehensively reflecting the local and global texture characteristics of the image. Morphological structure analysis focuses on the precise measurement and quantification of the geometric characteristics of specific biological structures in the image, such as collagen fiber bundles, including their width, the number of branch points, and their connection patterns. These parameters are of great significance for assessing the degree of liver fibrosis. Local color space transformation and data statistics can be understood as converting the image from the original color space such as RGB to a more suitable color space for analysis (such as HSV, Lab), and statistically analyzing the color distribution of local regions to obtain color distribution parameters that reflect information such as staining uniformity and nuclear staining intensity.

[0082] The extracted texture details, morphological structural parameters, and color distribution parameters are compared with a preset standard feature range. The purpose is to establish a benchmark to distinguish which features in the image represent genuine pathological changes and which are artifacts or distortions introduced by image processing operations (such as contrast enhancement and denoising). The preset standard feature range can be trained and learned from a large amount of labeled pathological image data, covering feature value ranges for different pathological stages and common artifact types. Through comparison, the degree of distortion in distorted feature regions can be quantified, providing more accurate input for subsequent reward calculations.

[0083] In practical applications, local region feature difference analysis is performed on intermediate processed images to identify regions in the image that show similarities in texture, morphology, and color distribution features. These regions may correspond to similar pathological features that require further fine differentiation.

[0084] This application's solution, by introducing multi-scale texture analysis, morphological structure analysis, and local color space transformation and data statistics, can comprehensively and meticulously capture various features of intermediate-processed images, laying a solid foundation for subsequent evaluation. Through the above technical solutions, this application can significantly improve the accuracy and robustness of reward computation networks in extracting multi-dimensional features from intermediate-processed images.

[0085] This application further proposes a context analysis method that focuses more on microstructure. By deeply extracting and fusing the microstructure features of images, it generates more complete and accurate context description information, thereby improving the adaptability and targeting of image processing.

[0086] The step of performing context analysis on the current state through the context analysis network in the above reinforcement learning model to generate context description information further includes: By using a context analysis network, multi-scale directional texture features of the input liver pathology image are extracted; based on the multi-scale directional texture features, a microstructure connection network corresponding to the input liver pathology image is constructed. By analyzing the connectivity components, directional consistency, and branch point density parameters of the microstructure connection network, the structural integrity score corresponding to the microstructure connection network is calculated. Based on the structural integrity score, combined with the local context information of the region where the microstructure connection network is located, the type of microstructure is identified and the degree of fuzziness of the microstructure is quantified. The obtained microstructure identification results and the degree of microstructure ambiguity are subjected to multi-dimensional feature fusion processing to integrate the microstructure feature information of the image; based on the fused microstructure feature information, the complete context description information is generated.

[0087] Specifically, extracting multi-scale directional texture features from input liver pathology images refers to capturing texture patterns in images at different spatial scales and directions using contextual analysis networks and techniques such as Gabor filter banks, wavelet transforms, or deep convolutional neural networks. These texture patterns can reflect subtle changes and directional arrangements of microstructures in liver tissue, such as collagen fibers and inflammatory cell infiltration, providing fundamental data for subsequent structural analysis.

[0088] The microstructure connection network constructed based on multi-scale directional texture features can be understood as using graph theory methods or morphological skeletonization techniques to transform the extracted texture features into a network structure. In this network, nodes can represent local feature points or key regions of the microstructure, while edges represent the connection relationships between these feature points or regions. The construction of this network aims to abstract and characterize the topological connections and spatial distribution characteristics of microstructures such as collagen fiber bundles and vascular networks in liver tissue.

[0089] In practical applications, the connectivity components, directional consistency, and branch point density parameters of the microstructural network are analyzed, and a structural integrity score is calculated. Specifically, the connectivity component parameter measures the number and size of independently connected microstructural regions in the network, reflecting the degree of dispersion or aggregation of the lesion area; the directional consistency parameter assesses the degree of orderliness or disorder in the arrangement of microstructures, such as the parallel arrangement or random distribution of collagen fibers in fibrotic tissue; and the branch point density parameter indicates the density of branch points in the network, which is closely related to pathological processes such as angiogenesis and fibrotic network remodeling.

[0090] Based on the analysis results of these parameters, a comprehensive structural integrity score is calculated using a pre-set weighted model, machine learning algorithm, or expert rule system. This score aims to quantify the health status or degree of disease of the microstructure.

[0091] Furthermore, based on the structural integrity score and combined with the local contextual information of the region where the microstructure connectivity network is located, the type of the microstructure is identified, and the degree of ambiguity of the microstructure is quantified. Local contextual information can include auxiliary features such as cell distribution, tissue density, and staining intensity around the microstructure. Type identification refers to classifying the identified microstructure into specific pathological types such as normal liver lobules, fibrous septa, and inflammatory cell clusters using a classifier (e.g., support vector machine, deep neural network) or a rule-based reasoning system. Simultaneously, the degree of ambiguity of the microstructure can be quantified using image sharpness indicators (e.g., gradient magnitude, spectral energy) or specialized image quality assessment models to reflect structural ambiguity caused by image quality or pathological changes.

[0092] Therefore, the obtained microstructure identification results and the degree of ambiguity of the microstructure are subjected to multi-dimensional feature fusion processing to integrate the microstructure feature information of the image. This fusion processing can integrate heterogeneous information, such as type labels (classification results) and ambiguity (quantified values), through feature vector concatenation, weighted averaging, attention mechanisms, or multimodal deep learning fusion networks to form a unified and comprehensive microstructure feature representation.

[0093] Finally, based on the fused microstructural features, a complete contextual description is generated. This contextual description can be a high-dimensional feature vector, a set of semantic labels, or a structured text description, used to comprehensively and accurately characterize the microscopic pathological state of the current input liver pathology image, providing precise guidance for the subsequent reinforcement learning model to generate customized image processing sequences.

[0094] The proposed solution overcomes the limitations of traditional contextual analysis, which may fail to capture key microscopic pathological features, by performing in-depth microscopic structural analysis on input liver pathology images. Through this technical solution, the contextual description information possesses high specificity and accuracy, more accurately reflecting the key features of microscopic lesions in liver pathology images. This enables reinforcement learning models to generate more targeted and adaptive image processing sequences.

[0095] refer to Figure 3 , Figure 3 This is a schematic diagram of a reinforcement learning-based intelligent image data processing system provided in an embodiment of the present invention, used to execute the above-described reinforcement learning-based intelligent image data processing method. The system includes: The state acquisition module is used to acquire the input liver pathology image and use the input liver pathology image as the current state of the reinforcement learning model. The context analysis module is used to perform context analysis on the current state through the context analysis network in the reinforcement learning model and generate context description information. The policy network module is used to generate customized image processing sequences based on the context description information through the policy network in the reinforcement learning model; An image processing module is used to process the input liver pathology image according to the image processing sequence to obtain an intermediate processed image; The reward calculation module is used to evaluate the intermediate processed image and calculate the reward signal through the reward calculation network in the reinforcement learning model; The parameter update module is used to update the network parameters of the policy network according to the reward signal; The iteration control module is used to cyclically call the policy network module, the image processing module, the reward calculation module, and the parameter update module until the reward signal meets the preset termination condition or reaches the preset maximum number of iterations, at which point the iteration stops and the current intermediate processed image is output as the final processed image.

[0096] This system employs a modular design to materialize the core functions of reinforcement learning methods, forming an end-to-end intelligent processing architecture. The state acquisition module transforms raw pathological images into inputs that the model can process, while the context analysis module performs deep image understanding, providing a basis for the policy network module to generate customized processing sequences. The image processing module executes specific image operations, the reward calculation module provides quantitative feedback on the processing results, and the parameter update module continuously optimizes the policy using feedback signals. The iterative control module coordinates the entire processing flow, ensuring that the system can adaptively optimize liver pathological images. This overcomes the subjectivity and inefficiency of traditional manual processing, as well as the insufficient generalization ability of existing intelligent systems when facing complex pathological images, significantly improving the automation, standardization, and accuracy of image processing.

[0097] The specific steps and working principle of the reinforcement learning-based intelligent image data processing method have been described in the above embodiments, and will not be repeated here. It should be emphasized that the intelligent image data processing system proposed in this application realizes the materialization and automated operation of the method by mapping each functional step in the above method to specific hardware or software modules.

[0098] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent image data processing based on reinforcement learning, characterized in that, include: The input liver pathology image is acquired and used as the current state of the reinforcement learning model. The context analysis network in the reinforcement learning model performs context analysis on the current state and generates context description information. The iterative process is repeated, which includes: generating a customized image processing sequence based on the context description information through the policy network in the reinforcement learning model; processing the input liver pathology image according to the image processing sequence to obtain an intermediate processed image; evaluating the intermediate processed image through the reward calculation network in the reinforcement learning model; calculating the corresponding reward signal; and updating the network parameters of the policy network according to the reward signal. The iteration stops when the reward signal meets the preset termination condition or the number of iterations reaches the preset maximum number of iterations, and the intermediate processed image generated in the current iteration is output as the final processed image. The reinforcement learning model is an end-to-end intelligent liver pathology image processing model that integrates the context analysis network, the policy network, and the reward calculation network. By iteratively optimizing the network parameters of the policy network, it achieves adaptive and customized intelligent data processing of the input liver pathology image, thus forming a complete reinforcement learning image data processing closed loop.

2. The intelligent image data processing method based on reinforcement learning according to claim 1, characterized in that, The step of performing context analysis on the current state through the context analysis network in the reinforcement learning model to generate context description information includes: Microstructural analysis, staining deviation analysis, and tissue background analysis were performed on the input liver pathology image to obtain microstructural feature information, first sub-context description information, and second sub-context description information, respectively. By fusing the microstructure feature information, the first sub-context description information, and the second sub-context description information, multi-dimensional fused context information is obtained, and the final context description information is generated based on the multi-dimensional fused context information.

3. The intelligent image data processing method based on reinforcement learning according to claim 2, characterized in that, The step of performing microstructural analysis on the input liver pathology image to obtain microstructural feature information includes: The context analysis network is used to extract multi-scale directional texture features from the input liver pathology image. Based on the multi-scale directional texture features, a microstructure connection network corresponding to the input liver pathology image is constructed. The connectivity parameters, orientation consistency parameters, and branch point density parameters of the microstructure connection network are analyzed, and the structural integrity score of the microstructure connection network is calculated based on the analysis results. Based on the structural integrity score, and combined with the local context information of the region where the microstructure is located corresponding to the microstructure connection network, the type of the microstructure is identified, and the image blur degree of the microstructure is quantified. Based on the identification results of the microstructure and the degree of ambiguity of the microstructure, the microstructure feature information is generated.

4. The intelligent image data processing method based on reinforcement learning according to claim 2, characterized in that, The step of performing staining bias analysis on the input liver pathology image to obtain the first sub-context description information includes: The context analysis network is used to segment and divide the input liver pathology image into multiple independent local regions. For each of the local regions, the color distribution features of the local region are extracted, and the cell density parameters and tissue arrangement regularity parameters corresponding to the local region are calculated. The color distribution characteristics are compared with a preset physiological color variation range benchmark to identify whether there is a staining deviation in the local area that exceeds the physiological range. Based on the cell density parameters and tissue arrangement regularity parameters of the local area, the identified staining deviations are weighted and corrected to distinguish between physiological color variations and staining artifacts. Based on the weighted correction of the staining deviation identification results, a first sub-context description information is generated to characterize the staining state of the image.

5. The intelligent image data processing method based on reinforcement learning according to claim 2, characterized in that, The step of performing tissue background analysis on the input liver pathology image to obtain the second sub-context description information includes: The context analysis network is used to segment and divide the input liver pathology image into multiple independent local regions. For each of the local regions, the tissue structure features of the local region are extracted, including cell density features and extracellular matrix distribution pattern features; Based on the tissue structure characteristics, the tissue type corresponding to the local area is identified, and the tissue type includes at least one of normal liver tissue, fatty degeneration area, and inflammatory infiltration area; Based on the identified organization type, a corresponding background feature evaluation strategy is selected from a preset strategy library. The background feature evaluation strategy predefines organizational background feature evaluation parameters and judgment thresholds corresponding to different organization types. Based on the selected background feature evaluation strategy, the organizational background features of the local area are quantitatively evaluated to identify whether there are abnormal features in the organizational background of the local area. Based on the anomaly assessment results of the described tissue background features, a second sub-contextual description is generated to characterize the tissue background state of the image.

6. The intelligent image data processing method based on reinforcement learning according to claim 1, characterized in that, The image processing sequence includes at least one of the following operations: adaptive histogram equalization, anisotropic diffusion filtering, morphological closing operation, contrast-limited adaptive histogram equalization, and Gaussian pyramid decomposition and reconstruction.

7. The intelligent image data processing method based on reinforcement learning according to claim 1, characterized in that, The step of evaluating the intermediate processed image and calculating the corresponding reward signal through the reward calculation network in the reinforcement learning model includes: The intermediate processed image is subjected to multi-dimensional feature extraction through the reward calculation network. The multi-dimensional features include texture features, morphological features, and color distribution features. The extracted multi-dimensional features are matched with a preset pathological feature template library and a preset artifact feature library to complete the identification and classification of key pathological features and artifact features in the intermediate processed image. Based on the identification and classification results of the key pathological features and the artifact features, the key pathological feature index and the artifact feature index are quantitatively calculated respectively. Based on the quantitative indicators of key pathological features and the quantitative indicators of artifact features, combined with the preset image evaluation rules, a comprehensive evaluation score is calculated, and the comprehensive evaluation score is output as the reward signal.

8. The intelligent image data processing method based on reinforcement learning according to claim 7, characterized in that, The step of extracting multi-dimensional features from the intermediate processed image through the reward calculation network includes: The reward calculation network is used to perform multi-scale texture analysis on the intermediate processed image to capture texture detail information of the intermediate processed image at different scales. Based on the acquired texture detail information, morphological structure analysis is performed on the intermediate processed image to identify and quantify the width parameters, branch point number parameters, and connection mode parameters of the collagen fiber bundles in the intermediate processed image. Combining the texture detail information with the morphological structure analysis results, local color space conversion and data statistics are performed on the intermediate processed image to obtain the color distribution parameters of each local region; The texture detail information, the morphological structure parameters, and the color distribution parameters are compared with the preset standard feature range to distinguish the feature distortion areas introduced by image processing and the real pathological change areas in the intermediate processed image. The preset standard feature range corresponds to the feature value range of different pathological stages and different artifact types. Based on the comparison results, the degree of distortion of the feature distortion region is quantified; The method further includes: performing local region feature difference analysis on the intermediate processed image to identify similar pathological feature regions with similar texture, morphology, and color distribution characteristics within the image; adaptively adjusting the local contrast enhancement parameters and spatial filtering parameters for the identified similar pathological feature regions to improve the boundary clarity between different types of pathological features; performing feature separation processing on the similar pathological feature regions after parameter adjustment to highlight the morphological differences between early diffuse collagen fibers and mature fiber intervals; and performing quantitative statistics on the image features after feature separation to generate highly distinguishable multi-dimensional features for subsequent comprehensive evaluation score calculation.

9. The intelligent image data processing method based on reinforcement learning according to claim 1, characterized in that, The step of performing context analysis on the current state through the context analysis network in the reinforcement learning model to generate context description information further includes: The context analysis network is used to extract multi-scale directional texture features from the input liver pathology image. Based on the multi-scale directional texture features, a microstructure connection network corresponding to the input liver pathology image is constructed. By analyzing the connectivity components, directional consistency, and branch point density parameters of the microstructure connection network, the structural integrity score corresponding to the microstructure connection network is calculated. Based on the structural integrity score, combined with the local context information of the region where the microstructure connection network is located, the type of microstructure is identified, and the degree of fuzziness of the microstructure is quantified. The obtained microstructure identification result and the degree of fuzziness of the microstructure are subjected to multi-dimensional feature fusion processing to integrate the image microstructure feature information; Based on the fused microstructural feature information, complete contextual description information is generated.

10. An intelligent image data processing system based on reinforcement learning, characterized in that, The system is used to perform the intelligent image data processing method based on reinforcement learning as described in any one of claims 1-9, the system comprising: The state acquisition module is used to acquire the input liver pathology image and use the input liver pathology image as the current state of the reinforcement learning model. The context analysis module is used to perform context analysis on the current state through the context analysis network in the reinforcement learning model and generate context description information. The policy network module is used to generate customized image processing sequences based on the context description information through the policy network in the reinforcement learning model; An image processing module is used to process the input liver pathology image according to the image processing sequence to obtain an intermediate processed image; The reward calculation module is used to evaluate the intermediate processed image and calculate the reward signal through the reward calculation network in the reinforcement learning model; The parameter update module is used to update the network parameters of the policy network according to the reward signal; The iteration control module is used to cyclically call the policy network module, the image processing module, the reward calculation module, and the parameter update module until the reward signal meets the preset termination condition or reaches the preset maximum number of iterations, at which point the iteration stops and the current intermediate processed image is output as the final processed image.

Images