Cerebral hemorrhage intelligent decision support system based on deep adaptive feature fusion

Abstract

The invention relates to the technical field, in particular to a cerebral hemorrhage intelligent decision support system based on deep adaptive feature fusion, which comprises a multi-modal data standardization module, a cross-modal feature fusion module, a focus segmentation calculation module, a hemorrhage type identification module and an illness state grading output module. According to the method, deep association between fusion features and bleeding types is mined through an attention mechanism, related feature indexes are converted into standardized scores by referring to clinical common scoring standards, model attention weight distribution is optimized by combining actual prognosis result deviation, the suitability of illness state grading and the clinical scoring standards is improved, and the probability of illness state grading is lowered. An accurate grading result and prognosis prediction reference are provided for clinical treatment decision and rehabilitation intervention; effective data are screened from multiple types of brain images according to image quality related indexes, clinical text key information extraction and association labeling are combined, data integrity and consistency verification is carried out at the same time, and the standardization degree of multi-source data is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision support technology for cerebral hemorrhage, and in particular to an intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion. Background Technology

[0002] The field of intelligent decision support technology for cerebral hemorrhage includes core technologies such as brain image processing, clinical text parsing, cross-modal feature integration, lesion segmentation calculation, hemorrhage type identification, and disease severity assessment. It is an important area of ​​cross-integration between medical image analysis and clinical decision support.

[0003] Among them, the intelligent decision support technology for cerebral hemorrhage based on deep adaptive feature fusion refers to the technical system formed to meet the data processing and clinical decision-making needs in the diagnosis and treatment of cerebral hemorrhage. This technology encompasses standardized processing of multi-source data, effective integration of cross-modal features, precise segmentation and volume measurement of lesions, differentiation of bleeding types with different causes, and disease grading and prognosis prediction.

[0004] Existing technologies often suffer from insufficient data standardization and inadequate exploitation of cross-modal feature complementarity when processing multi-source data from brain imaging and clinical texts. In lesion segmentation, the accuracy of identifying irregular hematoma edges is limited, and the volume measurement error is large. In hemorrhage type identification, the ability to distinguish rare causes is insufficient, and the model generalization is poor. In disease grading and prognosis prediction, there is a lack of in-depth correlation analysis of multi-dimensional features, and the adaptability with clinical scoring standards needs to be improved. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies and propose an intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion, the system comprising a multimodal data standardization module, a cross-modal feature fusion module, a lesion segmentation calculation module, a hemorrhage type identification module, and a disease severity grading output module; The multimodal data standardization module is used to process brain imaging and clinical text data; The cross-modal feature fusion module is used to integrate image and text features; The lesion segmentation calculation module is used to segment and calculate the volume of cerebral hemorrhage lesions. The hemorrhage type identification module is used to distinguish between different types of cerebral hemorrhage caused by different etiologies; The disease grading output module is used to generate disease grading and clinical scoring results.

[0007] Preferably, the multimodal data standardization module includes an image data cleaning submodule, a clinical text parsing submodule, and a dataset annotation submodule; The image data cleaning submodule collects multi-sequence image data from 16-slice / 64-slice CT and 1.5T / 3.0T MRI, filters data based on image resolution, lesion clarity, and anatomical structure integrity, and outputs valid image frames. The clinical text parsing submodule reads the patient's medical history, medication records, and laboratory test texts, extracts key information through natural language processing technology, and outputs text feature vectors. The dataset annotation submodule associates effective image frames with text feature vectors, annotates lesion geometric parameters, density / signal features, and clinical diagnostic results, and outputs a standardized multimodal dataset.

[0008] Preferably, the cross-modal feature fusion module includes an image feature extraction submodule, a text feature conversion submodule, and a dynamic weighted fusion submodule; The image feature extraction submodule uses a ResNet-101 backbone network to perform convolution operations on the effective frames of the image, extracting features related to hematoma morphology, density, and bleeding location, and outputting an image feature matrix. The text feature conversion submodule embeds and encodes text feature vectors, constructs clinical risk factor association data, and outputs a text feature matrix; The dynamic weighted fusion submodule uses an attention mechanism to assign weights to the image feature matrix and the text feature matrix and fuse them, outputting a fused feature tensor.

[0009] Preferably, the lesion segmentation calculation module includes a multi-scale feature processing submodule, an edge enhancement segmentation submodule, and a volume calculation submodule; The multi-scale feature processing submodule performs multi-scale input processing on the fused feature tensor, constructs an image pyramid, and outputs a multi-scale feature map. The edge enhancement segmentation submodule adopts a multi-scale recurrent fully convolutional U-Net architecture, which combines Canny edge detection and gradient loss function to perform segmentation operations on multi-scale feature maps and output lesion segmentation mask and hematoma edge coordinate set; The volume calculation submodule calculates the spatial proportion of the hematoma based on the lesion segmentation mask and the hematoma edge coordinate set, and outputs the hematoma volume value.

[0010] Preferably, the bleeding type identification module includes an attention feature optimization submodule, a transfer learning training submodule, and a type matching submodule; The attention feature optimization submodule is embedded after the ResNet-101 backbone network to process the fused feature tensor and output a class-specific attention map; The transfer learning training submodule uses a pre-trained model on the ImageNet dataset, combines it with a standardized multimodal dataset, fine-tunes the parameters using stochastic gradient descent, and outputs the trained model. The type matching submodule inputs the class-specific attention map into the trained model, compares it with the preset bleeding type features, and outputs the type recognition confidence and bleeding type code.

[0011] Preferably, the disease grading output module includes a multi-label feature association submodule, a clinical score mapping submodule, and a grading result generation submodule; The multi-label feature association submodule uses the ClassificationTransformer model to perform association operations on the fused feature tensor and the bleeding type code, and outputs the label association coefficient. The clinical scoring mapping submodule refers to the Graeb score and ICH scoring standards, converts the label correlation coefficient and hematoma volume value into corresponding scores, and outputs the clinical score value. The grading result generation submodule divides the range according to the clinical score value and outputs the disease grading code and the 30-day neurological function prognosis prediction value.

[0012] Preferably, the multimodal data standardization module further includes a data quality control submodule, which performs integrity and consistency checks on the standardized multimodal dataset, calculates the data missing ratio and annotation error ratio, and outputs a data quality control pass mark and a data missing rate value.

[0013] Preferably, the cross-modal feature fusion module further includes a feature dimensionality reduction submodule, which performs dimensionality reduction on the fused feature tensor using principal component analysis algorithm, retains principal components with a cumulative contribution rate ≥ 95%, eliminates redundant feature dimensions, and outputs the dimensionality-reduced feature matrix and principal component contribution rate.

[0014] Preferably, the bleeding type identification module further includes a reinforcement learning optimization submodule, which uses the type identification confidence as a reward signal, employs a temporal difference learning algorithm to adjust the parameter update step size of the trained model, and outputs parameter adjustment coefficients and model reward values.

[0015] Preferably, the disease grading output module further includes a reinforcement feedback submodule, which uses the deviation between the clinical score and the actual prognosis as a penalty signal, optimizes the attention weight allocation of the ClassificationTransformer model through a deep Q-network, and outputs an attention adjustment matrix and a prognosis deviation rate.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: 1. This invention filters effective data from multiple types of brain images according to image quality-related indicators, combines key information extraction and association annotation from clinical texts, and conducts data integrity and consistency verification to improve the standardization of multi-source data and provide high-quality data support for subsequent analysis.

[0017] 2. This invention extracts hematoma-related features from images using residual networks, embeds and encodes clinical risk factors in text, dynamically allocates weights through an attention mechanism to achieve cross-modal feature fusion, and then reduces dimensions to remove redundant information, fully exploring the complementarity of different modal data and enhancing the ability to distinguish complex cases.

[0018] 3. This invention constructs an image pyramid through multi-scale processing, combines edge detection technology with a specific loss function to optimize the segmentation architecture, enhances the ability to identify irregular hematoma edges, and calculates the spatial proportion of the hematoma based on pixel mapping according to the segmentation results, thereby improving the accuracy of lesion segmentation and volume measurement, and providing a reliable quantitative basis for disease assessment.

[0019] 4. This invention improves the ability to distinguish bleeding types caused by rare causes by focusing on the characteristic regions of bleeding from different etiologies after feature extraction, using a pre-trained model combined with clinical data to fine-tune parameters, and dynamically adjusting the model parameter update step size based on the identification confidence level. This enhances the model's generalization performance and improves the consistency and accuracy of bleeding type identification.

[0020] 5. This invention explores the deep correlation between fusion features and bleeding types through attention mechanisms, converts relevant feature indicators into standardized scores by referring to commonly used clinical scoring standards, and optimizes the model's attention weight allocation based on actual prognostic result deviations, thereby improving the adaptability of disease grading to clinical scoring standards and providing accurate grading results and prognostic prediction references for clinical treatment decisions and rehabilitation interventions. Attached Figure Description

[0021] Figure 1 This is a block diagram of the overall system flow of the present invention; Figure 2 This is a flowchart of a submodule of the multimodal data standardization module of the present invention; Figure 3 This is a flowchart of a submodule of the cross-modal feature fusion module of the present invention; Figure 4 This is a flowchart of a submodule of the lesion segmentation calculation module of the present invention; Figure 5 This is a flowchart of a submodule of the bleeding type identification module of the present invention; Figure 6 This is a flowchart of a submodule of the disease grading output module of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0023] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0024] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent.

[0025] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0026] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0027] Please see Figure 1 The present invention provides a technical solution: an intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion, the system including a multimodal data standardization module, a cross-modal feature fusion module, a lesion segmentation calculation module, a hemorrhage type identification module, and a disease severity grading output module; In the above, the multimodal data standardization module is used to process brain CT / MRI multi-sequence images and clinical text data to construct a standardized multimodal dataset, providing high-quality data support for subsequent analysis; In the above, the cross-modal feature fusion module is used to integrate image visual features and textual clinical risk factor features, and to make up for the limitations of single-modal analysis through a dynamic weighting mechanism, thereby enhancing the ability to distinguish complex cases; In the above, the lesion segmentation calculation module is used to achieve fine segmentation of cerebral hemorrhage lesions and accurate calculation of hematoma volume, providing a quantitative basis for disease assessment; the hemorrhage type identification module is used to distinguish cerebral hemorrhage types with various causes such as hypertensive hemorrhage, aneurysmal hemorrhage, and arteriovenous malformation, improving diagnostic consistency. The above-mentioned disease grading output module is used to generate disease grading, clinical scoring results and 30-day neurological function prognosis prediction values, providing intelligent support for clinical treatment decisions and rehabilitation interventions, and helping to promote the homogenization of primary healthcare.

[0028] Please see Figure 2 The multimodal data standardization module includes an image data cleaning submodule, a clinical text parsing submodule, and a dataset annotation submodule; The above-mentioned image data cleaning submodule collects multi-sequence image data (including plain / enhanced / functional imaging sequences) from 16-slice / 64-slice CT and 1.5T / 3.0T MRI. Based on indicators such as image resolution, lesion clarity, and integrity of cerebral vascular anatomy, it filters effective data and extracts effective image frames containing complete hemorrhage areas and surrounding tissues, covering different disease stages such as acute, subacute, and chronic phases. In the above, the clinical text parsing submodule reads text information such as patient medical history, medication records, laboratory tests, and underlying diseases, extracts key information and clinical risk factors through natural language processing technology, constructs an association network, and outputs text feature vectors; In the above, the dataset annotation submodule associates the effective frames of the image with the text feature vectors, annotates the geometric parameters of the lesion (major axis / minor axis / volume, distance from the ventricle, etc.), density / signal features (CT value range, signal intensity of each MRI sequence), and clinical diagnostic results (etiological classification, treatment plan, prognosis). The output is a standardized multimodal dataset covering more than 10,000 clinical cases, covering more than 95% of common clinical bleeding types.

[0029] Furthermore, the multimodal data standardization module also includes a data quality control submodule. This submodule performs integrity verification, consistency verification, and compliance review on the standardized multimodal dataset. It verifies the accuracy of annotations through an expert verification mechanism, calculates the proportion of missing data and the proportion of annotation errors, and outputs a data quality control pass mark, a data missing rate value, and an annotation error report to ensure that the dataset meets the requirements of clinical research and model training.

[0030] Please see Figure 3 The cross-modal feature fusion module includes an image feature extraction submodule, a text feature conversion submodule, and a dynamic weighted fusion submodule; In the above, the image feature extraction submodule adopts the ResNet-101 backbone network, solves the gradient vanishing problem of deep networks through residual connections, cancels fully connected layers and introduces global average pooling, performs convolution operation on effective frames of the image, extracts features related to hematoma morphology, density, bleeding location and surrounding tissue influence (edema / midline shift, etc.), and outputs the image feature matrix. In the above, the text feature transformation submodule embeds and encodes the text feature vectors, constructs clinical risk factor association data, clarifies the association pattern between medical history, medication history and bleeding risk, and outputs the text feature matrix; In the above, the dynamic weighted fusion submodule accurately captures the complementary relationship between image and text features through a cross-modal attention mechanism, dynamically assigns weights to the image feature matrix and the text feature matrix and fuses them, thereby enhancing the ability to distinguish complex cases (such as hypertensive hemorrhage and amyloid angiopathy hemorrhage) and outputting a fused feature tensor.

[0031] Furthermore, the cross-modal feature fusion module also includes a feature dimensionality reduction submodule. The feature dimensionality reduction submodule uses principal component analysis algorithm to reduce the dimensionality of the fused feature tensor, retains principal components with a cumulative contribution rate of ≥95%, eliminates redundant feature dimensions, reduces the computational load of the model and improves the generalization ability, and outputs the dimensionality-reduced feature matrix and the principal component contribution rate.

[0032] Please see Figure 4 The lesion segmentation calculation module includes a multi-scale feature processing submodule, an edge enhancement segmentation submodule, and a volume calculation submodule; In the above, the multi-scale feature processing submodule performs multi-scale input processing (128×128 / 256×256 / 512×512 pixels) on the fused feature tensor, constructs an image pyramid, captures the edge details of small hemorrhages through shallow paths, and fuses anatomical structures and semantic features of lesions through deep paths, outputting a multi-scale feature map. In the above, the edge enhancement segmentation submodule adopts a multi-scale recurrent fully convolutional U-Net architecture, combined with Canny edge detection preprocessing and gradient loss function, to enhance the segmentation ability of irregular hematoma boundaries, perform segmentation operations on multi-scale feature maps, and output lesion segmentation mask and hematoma edge coordinate set; In the above, the volume calculation submodule calculates the spatial proportion of the hematoma through pixel mapping based on the lesion segmentation mask and the hematoma edge coordinate set, and performs calibration in combination with the clinical gold standard to output an accurate hematoma volume value.

[0033] Please see Figure 5 The bleeding type identification module includes an attention feature optimization submodule, a transfer learning training submodule, and a type matching submodule; In the above, the attention feature optimization submodule embeds a class-specific residual attention (CSRA) module after the ResNet-101 backbone network. It generates class-specific attention maps through global average pooling, focuses on the feature regions of bleeding from different causes (such as focusing on the basal ganglia region for hypertensive bleeding), and outputs class-specific attention maps. In the above, the transfer learning training submodule uses a pre-trained model on the ImageNet dataset and combines it with a standardized multimodal dataset to fine-tune the parameters using stochastic gradient descent, quickly adapting to changes in the distribution of new case characteristics and outputting the trained model. In the above, the type matching submodule inputs the class-specific attention map into the trained model and compares it with the preset features of 12 types of bleeding, such as hypertension, aneurysm, cerebral arteriovenous malformation, amyloid angiopathy, and traumatic bleeding, and outputs the type recognition confidence and bleeding type code.

[0034] Furthermore, the bleeding type identification module also includes a reinforcement learning optimization submodule. The reinforcement learning optimization submodule uses the type identification confidence as the reward signal and adopts a temporal difference learning algorithm to dynamically adjust the parameter update step size of the training model, focusing on improving the identification ability of rare causes (such as dural arteriovenous fistula bleeding), and outputs parameter adjustment coefficients and model reward values.

[0035] Please see Figure 6 The disease grading output module includes a multi-label feature association submodule, a clinical score mapping submodule, and a grading result generation submodule. In the above, the multi-label feature association submodule adopts the ClassificationTransformer model, which learns the adaptive interaction between samples and labels through the attention mechanism, captures the long-distance association between bleeding site, volume and prognostic indicators, performs association operation on the fused feature tensor and bleeding type encoding, and outputs the label association coefficient. In the above, the clinical score mapping submodule refers to the Graeb score and ICH scoring standards, converts the label correlation coefficient, hematoma volume value and surrounding tissue influence index into corresponding scores, and outputs standardized clinical score values. In the above, the grading result generation submodule divides the mild, moderate and severe ranges according to the clinical score, and outputs the disease grading code, the 30-day neurological function prognosis prediction value and personalized treatment suggestions (such as blood pressure control target and surgical risk assessment).

[0036] Furthermore, the disease grading output module also includes a reinforcement feedback submodule. The reinforcement feedback submodule uses the deviation between the clinical score and the actual prognostic result as a penalty signal. It optimizes the attention weight allocation of the ClassificationTransformer model through a deep Q-network to reduce the prognostic assessment bias and outputs the attention adjustment matrix and prognostic bias rate to achieve continuous iterative optimization of the model.

[0037] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A brain hemorrhage intelligent decision support system based on deep adaptive feature fusion, characterized in that, The system includes a multimodal data standardization module, a cross-modal feature fusion module, a lesion segmentation calculation module, a bleeding type identification module, and a disease severity grading output module; The multimodal data standardization module is used to process brain imaging and clinical text data; The cross-modal feature fusion module is used to integrate image and text features; The lesion segmentation calculation module is used to segment and calculate the volume of cerebral hemorrhage lesions. The hemorrhage type identification module is used to distinguish between different types of cerebral hemorrhage caused by different etiologies; The disease grading output module is used to generate disease grading and clinical scoring results.

2. The intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion according to claim 1, characterized in that: The multimodal data standardization module includes an image data cleaning submodule, a clinical text parsing submodule, and a dataset annotation submodule; The image data cleaning submodule collects multi-sequence image data from 16-slice / 64-slice CT and 1.5T / 3.0T MRI, filters data based on image resolution, lesion clarity, and anatomical structure integrity, and outputs valid image frames. The clinical text parsing submodule reads the patient's medical history, medication records, and laboratory test texts, extracts key information through natural language processing technology, and outputs text feature vectors. The dataset annotation submodule associates effective image frames with text feature vectors, annotates lesion geometric parameters, density / signal features, and clinical diagnostic results, and outputs a standardized multimodal dataset.

3. The intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion according to claim 1, characterized in that: The cross-modal feature fusion module includes an image feature extraction submodule, a text feature conversion submodule, and a dynamic weighted fusion submodule; The image feature extraction submodule uses a ResNet-101 backbone network to perform convolution operations on the effective frames of the image, extracting features related to hematoma morphology, density, and bleeding location, and outputting an image feature matrix. The text feature conversion submodule embeds and encodes text feature vectors, constructs clinical risk factor association data, and outputs a text feature matrix; The dynamic weighted fusion submodule uses an attention mechanism to assign weights to the image feature matrix and the text feature matrix and fuse them, outputting a fused feature tensor.

4. The intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion according to claim 1, characterized in that: The lesion segmentation calculation module includes a multi-scale feature processing submodule, an edge enhancement segmentation submodule, and a volume calculation submodule; The multi-scale feature processing submodule performs multi-scale input processing on the fused feature tensor, constructs an image pyramid, and outputs a multi-scale feature map. The edge enhancement segmentation submodule adopts a multi-scale recurrent fully convolutional U-Net architecture, which combines Canny edge detection and gradient loss function to perform segmentation operations on multi-scale feature maps and output lesion segmentation mask and hematoma edge coordinate set; The volume calculation submodule calculates the spatial proportion of the hematoma based on the lesion segmentation mask and the hematoma edge coordinate set, and outputs the hematoma volume value.

5. The intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion according to claim 1, characterized in that: The bleeding type identification module includes an attention feature optimization submodule, a transfer learning training submodule, and a type matching submodule; The attention feature optimization submodule is embedded after the ResNet-101 backbone network to process the fused feature tensor and output a class-specific attention map; The transfer learning training submodule uses a pre-trained model on the ImageNet dataset, combines it with a standardized multimodal dataset, fine-tunes the parameters using stochastic gradient descent, and outputs the trained model. The type matching submodule inputs the class-specific attention map into the trained model, compares it with the preset bleeding type features, and outputs the type recognition confidence and bleeding type code.

6. The intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion according to claim 1, characterized in that: The disease grading output module includes a multi-label feature association submodule, a clinical score mapping submodule, and a grading result generation submodule. The multi-label feature association submodule uses the ClassificationTransformer model to perform association operations on the fused feature tensor and the bleeding type code, and outputs the label association coefficient. The clinical scoring mapping submodule refers to the Graeb score and ICH scoring standards, converts the label correlation coefficient and hematoma volume value into corresponding scores, and outputs the clinical score value. The grading result generation submodule divides the range according to the clinical score value and outputs the disease grading code and the 30-day neurological function prognosis prediction value.

7. The intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion according to claim 2, characterized in that: The multimodal data standardization module also includes a data quality control submodule, which performs integrity and consistency checks on the standardized multimodal dataset, calculates the data missing ratio and annotation error ratio, and outputs the data quality control pass mark and the data missing rate value.

8. The intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion according to claim 3, characterized in that: The cross-modal feature fusion module also includes a feature dimensionality reduction submodule. The feature dimensionality reduction submodule uses principal component analysis algorithm to reduce the dimensionality of the fused feature tensor, retains principal components with a cumulative contribution rate of ≥95%, removes redundant feature dimensions, and outputs the dimensionality-reduced feature matrix and the principal component contribution rate.

9. The intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion according to claim 5, characterized in that: The bleeding type identification module also includes a reinforcement learning optimization submodule. The reinforcement learning optimization submodule uses the type identification confidence as the reward signal, adopts a temporal difference learning algorithm to adjust the parameter update step size of the trained model, and outputs the parameter adjustment coefficient and the model reward value.

10. The intelligent decision support system for cerebral hemorrhage based on deep adaptive feature fusion according to claim 6, characterized in that: The disease grading output module also includes a reinforcement feedback submodule. The reinforcement feedback submodule uses the deviation between the clinical score and the actual prognosis as a penalty signal, optimizes the attention weight allocation of the ClassificationTransformer model through a deep Q-network, and outputs the attention adjustment matrix and the prognosis deviation rate.

Images