AI auxiliary remote consultation system for testicular nucleoprotein cancer

Abstract

The invention discloses an AI auxiliary remote consultation system for testicular nucleoprotein carcinoma, and relates to the technical field of medical information, and the system comprises a multi-modal data fusion processing module which is used for receiving and preprocessing patient multi-modal medical data from different medical institutions; the dynamic feature extraction and tumor evolution modeling module is used for extracting dynamic pathological features related to the testicular nucleoprotein cancer from the fused multi-dimensional feature tensor; the core of the expert diagnosis logic deep simulation module is a deep reinforcement learning network which is pre-trained and continuously learns online; and the real-time interaction and decision synchronization module is used for coordinating data streams and control signals among the modules in the system and managing real-time communication with a remote terminal user. The system solves the problems that an existing system is insufficient in testicular nucleoprotein cancer multi-modal data fusion capability, poor in rare case generalization and low in real-time decision-making efficiency, and the diagnosis accuracy and the remote cooperation efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical information technology, and in particular to an AI-assisted remote consultation system for testicular nuclear protein cancer. BACKGROUND

[0002] The deep integration of artificial intelligence and telemedicine technology is reshaping the pattern of modern medical services, providing unprecedented technical support and decision-making assistance for clinical diagnosis and treatment. The AI-based auxiliary diagnosis system significantly improves the efficiency and accessibility of medical services by efficiently integrating multi-modal medical images, massive clinical data, and scarce expert resources, becoming a key technical support for promoting global medical resource equalization and precision development.

[0003] Among them, the remote consultation system for rare tumor diseases is an important application direction of artificial intelligence in the medical field. This kind of system aims to analyze the multi-modal medical data of patients through intelligent algorithms, and provide professional diagnosis suggestions and treatment plan recommendations for clinicians distributed in different regions, so as to make up for the imbalance of medical resources among regions.

[0004] The remote consultation system in the prior art faces significant challenges in dealing with testicular nuclear protein cancer, a rare malignant tumor. The system lacks the ability to analyze multi-modal medical data, making it difficult to accurately identify the subtle features and pathological evolution of the tumor. Traditional diagnosis models rely on limited training samples, resulting in poor generalization performance for rare cases, which cannot meet the strict requirements of clinical diagnosis accuracy. In addition, the existing system lacks the ability to learn the logic of expert diagnosis, making it difficult to provide reliable decision support in complex condition judgment, and the data synchronization and interaction efficiency in real-time consultation process is low, which seriously affects the timeliness of diagnosis and treatment. SUMMARY

[0005] The purpose of the present application is to provide an AI-assisted remote consultation system for testicular nuclear protein cancer, to solve the problems of insufficient multi-modal medical data fusion analysis ability, poor generalization performance for rare cases, lack of expert diagnosis logic deep learning ability, and low real-time interaction efficiency in the prior art.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] An AI-assisted remote consultation system for testicular nuclear protein cancer, comprising a multi-modal data fusion processing module, a dynamic feature extraction and tumor evolution modeling module, an expert diagnosis logic deep simulation module, and a real-time interaction and decision synchronization module. The multi-modal data fusion processing module is used to receive and preprocess multi-modal medical data of patients from different medical institutions, including medical image data, pathological section data, genomics data, and clinical text data. The module standardizes the original data through a preset data cleaning process, and uses a feature weighting fusion algorithm based on attention mechanism to align and integrate data from different modalities in a unified high-dimensional feature space, generating a fused multi-dimensional feature tensor. The dynamic feature extraction and tumor evolution modeling module is connected to the output end of the multi-modal data fusion processing module, and is used to extract dynamic pathological features related to testicular nuclear protein cancer from the fused multi-dimensional feature tensor. The module has a tumor evolution model based on a hybrid architecture of time convolution network and long short-term memory network, which takes the extracted dynamic pathological feature sequence as input, learns the evolution law of the tumor in the time dimension through multiple layers of neural network, and outputs the prediction probability distribution of the tumor state at the next three time points. The expert diagnosis logic deep simulation module is further connected to the dynamic feature extraction and tumor evolution modeling module, and its core is a pre-trained and continuously online learning deep reinforcement learning network. The network takes the prediction probability distribution output by the tumor evolution model and the multi-dimensional feature tensor of the current patient as state input, simulates the diagnosis reasoning process of human tumor experts under complex conditions, and outputs comprehensive diagnosis suggestions including tumor staging, treatment scheme recommendation priority ranking, and prognosis risk assessment. The real-time interaction and decision synchronization module is responsible for coordinating the data flow and control signals between the modules in the system, and managing real-time communication with remote terminal users. The module uses a bidirectional data channel based on Web real-time communication protocol to ensure that diagnosis suggestions, medical image annotation information, and user feedback data are transmitted end-to-end within 50 milliseconds. At the same time, the module integrates a version control mechanism to timestamp and logically associate store all generated diagnosis decisions and their basis, ensuring the consistency and traceability of decision history in multi-specialist collaborative consultation scenarios.

[0008] Further, the data cleaning process in the multi-modal data fusion processing module specifically includes three sub-steps of outlier detection and correction, data format unification, and missing value imputation. Outlier detection uses an unsupervised learning method based on the isolation forest algorithm to identify and correct noise artifacts in medical images and logical contradiction items in pathological texts. Data format unification resamples medical image data from different sources to 1 millimeter isotropic resolution, and converts all text data to a unified standardized medical terminology code. Missing value imputation uses a generative adversarial network model to generate reasonable synthetic data to fill in missing fields according to the existing data distribution.

[0009] Furthermore, the execution process of the attention-based feature weighted fusion algorithm is as follows: First, a feature encoder is trained for each modality of data, mapping the original data to a 512-dimensional common latent space. Medical image data is encoded using a deep residual network, pathological slide data using a convolutional neural network, genomics data using a fully connected neural network, and clinical text data using a bidirectional Transformer encoder. Then, the importance weight of each modality's feature vector relative to the current diagnostic task is calculated. This weight is dynamically generated by a trainable multilayer perceptron based on the feature vector itself and its contextual information. Finally, the feature vectors of all modalities are weighted and summed according to their corresponding importance weights to generate the final multidimensional feature tensor.

[0010] Furthermore, the tumor evolution model in the dynamic feature extraction and tumor evolution modeling module adopts an encoder-decoder structure. The encoder consists of a 4-layer temporal convolutional network, responsible for extracting high-level temporal patterns from the input feature sequence. The decoder consists of a 2-layer long short-term memory network, responsible for autoregressively generating predictions of tumor status at future time points based on the context vector output by the encoder. During the training phase, the model uses a weighted combination of the mean squared error loss function and the KL divergence loss function as the optimization objective to ensure that the prediction results are both accurate and conform to the medical prior distribution.

[0011] Furthermore, the deep reinforcement learning network in the deep simulation module of the expert diagnostic logic adopts an actor-critic architecture. The actor network is responsible for outputting a diagnostic action policy based on the current system state, which is a probability distribution over all possible diagnostic actions. The critic network is responsible for evaluating the value function of the current state, i.e., the long-term expected cumulative reward for executing the actor network's policy. The network's reward signal consists of three parts: a consistency reward between the diagnostic action and the real expert's diagnostic record, a delayed reward for the diagnostic result being verified as correct by subsequent clinical follow-up, and a penalty for the uncertainty of the diagnostic decision. The network is trained using a proximal policy optimization algorithm to stably learn the optimal diagnostic policy.

[0012] Furthermore, the version control mechanism in the real-time interaction and decision synchronization module operates as follows: The system assigns a globally unique identifier to each diagnostic session. Each intermediate diagnostic conclusion or final decision generated within that session is considered a version node, with nodes connected by directional edges to represent the derivation path of the diagnostic logic. Each node stores the decision content, decision timestamp, a list of characteristic sources for the decision basis, and the decision confidence level. When multiple experts consult on the same case, the system detects and highlights the points of divergence between the decision paths of different experts and automatically initiates an evidence-based conflict resolution process.

[0013] Furthermore, the system also includes a distributed model update and knowledge distillation subsystem. Deployed on a cloud server cluster, this subsystem periodically collects anonymized diagnostic data and expert feedback from various participating medical institutions. Using this new data, the core AI model in the system is incrementally updated through a federated learning framework, while avoiding centralized storage of raw patient data. The knowledge distillation component compresses the complex policy network learned from the deep simulation module of expert diagnostic logic into a lightweight student network. This student network can be deployed on resource-constrained edge computing devices to support preliminary diagnosis in offline environments.

[0014] Furthermore, the system's security and privacy protection layer is integrated into all data communication and storage stages. All data leaving the user's end is encrypted using encryption algorithms that comply with national cryptographic management standards before transmission. Patient identification information stored in the system database is replaced with irreversible hash values, while sensitive medical data is segmented into multiple data fragments and stored separately on different physical servers, and can only be reassembled and accessed with the patient's explicit authorization.

[0015] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0016] This invention achieves deep alignment and integration of medical imaging, pathology, genomic and clinical text data related to testicular nucleoprotein carcinoma by constructing a multimodal data fusion processing module, effectively overcoming the limitations of single data source analysis and providing a high-quality data foundation for subsequent accurate analysis.

[0017] This invention utilizes a hybrid neural network architecture to capture the temporal evolution patterns of tumors through a dynamic feature extraction and tumor evolution modeling module, significantly improving the predictive ability for the future development of this rare cancer and assisting doctors in proactive intervention. The expert diagnostic logic deep simulation module simulates the diagnostic reasoning patterns of top experts through deep reinforcement learning technology, enabling the system to provide highly valuable diagnostic suggestions even when faced with complex and atypical rare cases, greatly compensating for the lack of experience in the diagnosis and treatment of testicular nucleoprotein cancer in primary healthcare institutions.

[0018] This invention ensures the efficiency and collaboration of the consultation process through a real-time interaction and decision synchronization module. Its built-in version control mechanism guarantees the transparency and auditability of the decision-making process, which is conducive to knowledge sharing and decision optimization among multidisciplinary teams. The distributed updates and rigorous security design of the system as a whole not only ensure the continuous evolution and adaptability of the AI ​​model, but also ensure the security of patient privacy data throughout the consultation process, providing a solid technical guarantee for large-scale clinical application. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical architecture of the AI-assisted remote consultation system for testicular nucleoprotein carcinoma proposed in this invention.

[0020] Figure 2 This is a schematic diagram of the core principle framework of the feature weighted fusion algorithm based on the attention mechanism in this invention. Detailed Implementation

[0021] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0022] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0023] In the embodiments of the present invention, the same reference numerals denote the same components, and for the sake of brevity, detailed descriptions of the same components are omitted in different embodiments. It should be understood that the thickness, length, width, and other dimensions of various components in the embodiments of the present invention shown in the accompanying drawings, as well as the overall thickness, length, width, and other dimensions of the integrated device, are merely illustrative and should not constitute any limitation on the present invention; the term "multiple" in the present invention refers to two or more (including two).

[0024] Example 1

[0025] Please refer to the attached document. Figure 1This embodiment details the specific technical implementation of an AI-assisted remote consultation system for testicular nucleoprotein carcinoma. The system's core consists of a multimodal data fusion processing module, a dynamic feature extraction and tumor evolution modeling module, an expert diagnostic logic deep simulation module, and a real-time interaction and decision synchronization module. These modules are tightly coupled and work collaboratively through a high-speed data bus and control logic circuits.

[0026] The multimodal data fusion processing module serves as the system's data entry point, responsible for connecting to heterogeneous medical data sources from different medical institutions. This module incorporates a data cleaning unit, whose execution process strictly follows three sub-steps: outlier detection and correction, data format standardization, and missing value imputation. In the outlier detection and correction sub-step, the system invokes the Isolation Forest algorithm engine to perform unsupervised learning analysis on the input data. This engine first constructs multiple isolation tree data structures, with each tree randomly selecting features and split points to isolate data instances one by one. Noise artifacts in medical image data manifest as extreme outliers in pixel intensity values, while logical contradictions in pathological text data are identified as anomalous patterns using a semantic consistency verification model. The system calculates an anomaly score for each data instance, based on the path length from the root node to a leaf node within the isolation tree. When the anomaly score exceeds a preset threshold of 0.95, the system automatically triggers a correction procedure. For medical image noise, a filtering algorithm based on neighborhood pixel weighted averaging is used for smoothing; for textual logical contradictions, semantic reconstruction and replacement are performed based on a standardized medical terminology knowledge base. In the data format standardization sub-step, the system initializes the format conversion engine. For medical image data, this engine first analyzes its original resolution and spatial dimension parameters, then calls a 3D resampling algorithm to interpolate all voxel data to a 1-millimeter isotropic resolution. This process uses trilinear interpolation to ensure that the geometric features of the image are not distorted. For text data, the format conversion engine integrates a natural language processing pipeline, using entity recognition and normalization components to convert unstructured clinical descriptions into a coding system conforming to the International Classification of Diseases, 10th Revision (ICD-10), while simultaneously converting all time information to Coordinated Universal Time (UTC) timestamps. In the missing value imputation step, the system activates a generative adversarial network (GAN) model. This model consists of a generator and a discriminator. The generator receives complete data samples as training input, learns their latent distribution characteristics, and then generates reasonable synthetic values ​​for missing fields based on existing data fields. The discriminator is responsible for distinguishing between real complete samples and generated samples, continuously optimizing the generator's data generation quality through adversarial training. The system selects different generator architectures based on the data type of the missing field: a fully connected neural network generator is used for continuous numerical missing values, while a conditional generative adversarial network architecture is used for categorical field missing values. After the data cleaning unit completes the processing, the data enters the feature weighted fusion algorithm unit based on an attention mechanism.

[0027] Please refer to the attached document.Figure 2 The attention-based feature weighted fusion algorithm unit is responsible for mapping and integrating multimodal data to a unified feature space. This unit deploys four independent feature encoders, processing medical image data, pathological slide data, genomics data, and clinical text data respectively. The medical image data feature encoder is implemented using a deep residual network. Its input is a preprocessed 3D medical image tensor. The network structure contains five residual blocks, each consisting of two 3x3 convolutional layers, a batch normalization layer, and a modified linear unit activation function. Finally, a global average pooling layer outputs a 512-dimensional feature vector. The pathological slide data feature encoder is implemented using a convolutional neural network. Its input is a high-resolution digital pathological image. The network structure contains eight convolutional layers and three max-pooling layers. The convolutional kernel size decreases layer by layer from 7x7 to 3x3, and a fully connected layer outputs a 512-dimensional feature vector. The genomics data feature encoder is implemented using a fully connected neural network. Its input is a numerical vector of gene expression profiles or mutation profiles. The network structure contains three hidden layers, with 1024, 512, and 256 neurons per layer, respectively, ultimately outputting a 512-dimensional feature vector. The clinical text data feature encoder is implemented using a bidirectional Transformer encoder. Its input is a text sequence after word segmentation and embedding representation. The encoder has six layers, with eight self-attention heads per layer, ultimately outputting a 512-dimensional feature vector through a sequence pooling layer. After all feature encoders output, the system calls the attention weight calculation module. This module is implemented using a three-layer perceptron. Its input is the concatenated multimodal feature vector and its contextual metadata. The number of neurons in the hidden layers is 256 and 128, respectively. The output layer generates attention weight values ​​corresponding to the four modalities through a Softmax function. These weight values ​​dynamically reflect the importance of each modality of data to the current diagnostic task. The final fusion process is expressed by the following formula:

[0028]

[0029] in, Indicates the first Attention weights for each modality Indicates the first The system outputs a 512-dimensional feature vector from the feature encoder of each modality. The system uses a backpropagation algorithm to jointly optimize the parameters of the feature encoder and the attention weight calculation module, ensuring that the fused multidimensional feature tensor retains, to the greatest extent possible, cross-modal information valuable for the diagnosis of testicular nucleoprotein carcinoma. This multidimensional feature tensor serves as an intermediate representation of the system and is transmitted to the dynamic feature extraction and tumor evolution modeling module.

[0030] The dynamic feature extraction and tumor evolution modeling module receives multidimensional feature tensors from upstream sources and performs dynamic pathological feature extraction and tumor evolution prediction functions. This module first performs sliding window processing on the input feature sequence through a temporal feature extraction unit, with the window size set to five time points and the stride set to one time point. The feature sequence within each window is fed into the tumor evolution model for deep analysis. The tumor evolution model adopts an encoder-decoder hybrid architecture. The encoder consists of four stacked temporal convolutional networks. Each temporal convolutional network layer includes one-dimensional causal convolution operations, weight normalization techniques, and gated activation units. The convolutional kernel size is set to three along the time dimension, and the number of filters increases layer by layer to 64, 128, 256, and 512. The encoder outputs a condensed context vector that captures high-level temporal dependencies in the input feature sequence. The decoder consists of two layers of long short-term memory networks, with 256 hidden units per layer. The decoder uses the context vector output by the encoder as the initial hidden state and recursively generates tumor state predictions for the next three time points in an autoregressive manner. The predicted output at each time point is a probability distribution vector, with its dimensions corresponding to different clinical states of the tumor, such as stable, progressive, and remission. The tumor evolution model is trained using a combined loss function, which is a weighted average of a mean squared error (MSE) term and a KL divergence term. The MSE term measures the numerical difference between the predicted probability distribution and the true label distribution, while the KL divergence term constrains the consistency between the predicted distribution and the statistical distribution in the medical prior knowledge base. The weight ratio of the two loss terms is determined to be 7:3 through grid search. The model updates its parameters using stochastic gradient descent, with the learning rate initially set to 0.01 and dynamically adjusted using cosine annealing.

[0031] The expert diagnostic logic deep simulation module is connected to the tumor evolution modeling module, and its core component is a deep reinforcement learning network. This network adopts an actor-commentator architecture, receiving the predicted probability distribution output by the tumor evolution model and the current patient's multidimensional feature tensor as its state representation. The state representation first passes through a state encoding layer, which consists of two fully connected layers, mapping heterogeneous state information into a unified 256-dimensional state vector. The actor network receives the state vector as input, and its network structure contains three hidden layers with 512, 256, and 128 neurons respectively. The output layer generates a diagnostic action policy probability distribution through a Softmax function. The diagnostic action space covers discrete decision options such as tumor staging, treatment plan recommendation, and prognostic risk assessment. For example, tumor staging is divided into stages I, II, III, and IV, and treatment plans are prioritized as surgical resection, radiotherapy, chemotherapy, targeted therapy, and immunotherapy. The commentator network also receives the state vector as input, and its network structure is similar to the actor network, but its output layer is a single scalar value representing the value function estimate of the current state. The value function estimate represents the long-term expected cumulative reward obtained by performing diagnostic actions following the actor network policy, starting from the current state. The reward function design integrates three factors: the consistency reward between the diagnostic action and the real expert's diagnostic record is calculated through exact matching; a successful match results in a positive reward of one, otherwise zero. The delayed reward, positive two, is issued after the follow-up data arrives, provided the diagnostic result is verified as correct through subsequent clinical follow-up. The penalty for uncertainty in the diagnostic decision is calculated based on the entropy of the policy distribution; a higher entropy indicates greater uncertainty and a higher penalty. Network training employs a proximal policy optimization algorithm, which ensures the stability of policy updates through importance sampling and pruning mechanisms. During training, the system periodically samples batch trajectory data from the experience replay buffer, calculates the advantage function estimate, and updates the parameters of the actor and critic networks. The learning rate is set to 0.0001, and the pruning parameter epsilon is set to 0.2.

[0032] The real-time interaction and decision synchronization module acts as the system coordination center, managing internal data flow and external communication. This module integrates a real-time communication engine, employing a Web real-time communication protocol to establish a bidirectional data channel with remote terminal users. The communication engine maintains multiple virtual channels to transmit diagnostic suggestions, medical image annotation information, user feedback data, and control commands. The system continuously measures end-to-end transmission latency through a service quality monitoring unit, ensuring all data packets are transmitted within fifty milliseconds. When network congestion causes latency to exceed a threshold, the system automatically activates data compression algorithms and differential update mechanisms, prioritizing the real-time performance of critical diagnostic information. This module also incorporates a version control mechanism, assigning a 128-bit globally unique identifier to each diagnostic session. The version control mechanism models each diagnostic conclusion or decision as a version node. The node data structure includes a decision content field, a decision timestamp field, a list of feature sources for the decision basis field, and a decision confidence field. Nodes are connected by directed edges, forming a diagnostic logic derivation graph. When the system generates a new diagnostic decision, the version control mechanism automatically creates a new node and establishes a derivation edge with its parent node. The derivation edge stores the inference rules or model output information used. In multi-expert collaborative consultation scenarios, the system creates an independent decision branch for each expert. When a divergence is detected between different expert decision paths, the system triggers a conflict detection algorithm. This algorithm compares the output differences of each branch at key decision nodes. When the difference exceeds a preset threshold of 0.8, the system automatically highlights the divergence point and initiates an evidence-driven conflict resolution process. The conflict resolution process first collects the feature evidence and intermediate model results relied upon by each decision branch, and then generates consistent decision recommendations through a weighted voting algorithm or a confidence-based weighted fusion algorithm. All version nodes and their relationships are persistently stored in a graph database, supporting historical decision tracing and auditing by multiple dimensions such as time range, decision type, and expert identity.

[0033] The system also integrates a distributed model update and knowledge distillation subsystem, deployed on a cloud server cluster. The model update component employs a federated learning framework, periodically collecting anonymized diagnostic data and expert feedback from participating medical institutions. Each participating institution locally calculates the gradient of its model parameters using the new data, then uploads the encrypted gradient to the aggregation server. The aggregation server aggregates the gradients using secure multi-party computation technology, generating a global model update. The updated model parameters are distributed back to each institution, achieving collaborative model evolution without exposing the original data. The knowledge distillation component compresses the complex policy network in the deep simulation module of the expert diagnostic logic into a lightweight student network. The distillation process adopts a teacher-student learning paradigm, using the original policy network as the teacher network and its output probability distribution as the soft objective. The student network uses a simplified fully connected architecture, containing only two hidden layers with sixty-four neurons per layer. The distillation loss function is a combination of cross-entropy loss and mean squared error loss, ensuring that the student network can both mimic the decision distribution of the teacher network and maintain prediction accuracy. After training, the student network model is reduced to one-twentieth the size of the original network and can be deployed to edge computing devices to perform preliminary diagnostic inference in an offline environment.

[0034] The system's security and privacy protection layers are implemented throughout all data processing stages. During data transmission, the system uses the SM4 block cipher algorithm approved by the State Cryptography Administration to encrypt outbound data. The encryption key is dynamically negotiated through an elliptic curve-based key exchange protocol and updated with each session. During data storage, patient identification information is converted into a 256-bit irreversible hash value using the SM3 cryptographic hash algorithm. Sensitive medical data is segmented into multiple data fragments, with a secret sharing scheme used for fragmentation. Each fragment is stored separately on a different physical server node. Data reconstruction access requires two conditions: first, obtaining an access token authorized by the patient through digital signature; and second, collecting a sufficient number of data fragments from at least three server nodes to perform reconstruction calculations. All access operations are recorded in a security audit log. Log entries include operation time, operation type, user identifier, and resource identifier. The log file is stored tamper-proofly using blockchain technology.

[0035] This embodiment, through the detailed modular design and engineering implementation described above, constructs a highly professional, automated, and collaborative AI-assisted remote consultation system. The system begins with multimodal data fusion, proceeds through dynamic feature extraction and tumor evolution modeling, then generates comprehensive diagnostic suggestions through expert diagnostic logic simulation, and finally delivers the service to the end user via an efficient real-time interaction and decision synchronization mechanism. The entire process operates within a strict security and privacy protection framework, and continuous optimization and widespread deployment are achieved through distributed learning and knowledge distillation technologies, ultimately providing strong technical support for the precision diagnosis and treatment of testicular nucleoprotein carcinoma, a rare disease.

[0036] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An AI-assisted remote consultation system for testicular nucleocarcinoma, characterized in that, include: The multimodal data fusion processing module is used to receive and preprocess multimodal medical data of patients from different medical institutions. The multimodal data fusion processing module standardizes the raw data through a preset data cleaning process, and uses a feature weighted fusion algorithm based on attention mechanism to align and integrate the data from different modalities in a unified high-dimensional feature space to generate a fused multidimensional feature tensor. The dynamic feature extraction and tumor evolution modeling module is used to extract dynamic pathological features related to testicular nucleoprotein carcinoma from the fused multidimensional feature tensor. The dynamic feature extraction and tumor evolution modeling module has a built-in tumor evolution model based on a hybrid architecture of temporal convolutional network and long short-term memory network, which outputs the probability distribution of tumor status prediction for the next three time points. The expert diagnostic logic deep simulation module is based on a deep reinforcement learning network that is pre-trained and continuously learns online. The deep reinforcement learning network takes the predicted probability distribution output by the tumor evolution model and the multidimensional feature tensor of the current patient as state inputs to simulate the diagnostic reasoning process of human oncology experts under complex conditions and output comprehensive diagnostic suggestions. The real-time interaction and decision synchronization module is used to coordinate the data flow and control signals between various modules within the system, and to manage real-time communication with remote terminal users. The real-time interaction and decision synchronization module adopts a bidirectional data channel based on the Web real-time communication protocol. At the same time, the real-time interaction and decision synchronization module integrates a version control mechanism to timestamp and logically associate all generated diagnostic decisions and their basis.

2. The AI-assisted remote consultation system for testicular nucleocarcinoma according to claim 1, characterized in that, The data cleaning process in the multimodal data fusion processing module includes three sub-steps: outlier detection and correction, data format standardization, and missing value imputation. The outlier detection uses an unsupervised learning method based on the isolated forest algorithm to identify and correct noise artifacts in medical images and logical contradictions in pathological texts. The data format standardization resamples medical image data from different sources to a 1 mm isotropic resolution and converts all text data into a unified standardized medical terminology code. The missing value imputation uses a generative adversarial network model to generate reasonable synthetic data based on the existing data distribution to fill in missing fields.

3. The AI-assisted remote consultation system for testicular nucleocarcinoma according to claim 1, characterized in that, The execution process of the attention-based feature weighted fusion algorithm is as follows: a feature encoder is trained for each modality of data, mapping the original data to a 512-dimensional common latent space; medical image data is encoded using a deep residual network, pathological slide data is encoded using a convolutional neural network, genomics data is encoded using a fully connected neural network, and clinical text data is encoded using a bidirectional Transformer encoder; the importance weight of each modality feature vector relative to the current diagnostic task is calculated, and the importance weight is dynamically generated by a trainable multilayer perceptron based on the feature vector itself and its context information; the feature vectors of all modalities are weighted and summed according to their corresponding importance weights to generate the final multidimensional feature tensor.

4. The AI-assisted remote consultation system for testicular nucleocarcinoma according to claim 1, characterized in that, The tumor evolution model adopts an encoder-decoder structure; the encoder consists of a 4-layer temporal convolutional network, which is responsible for extracting high-level temporal patterns from the input feature sequence; the decoder consists of a 2-layer long short-term memory network, which is responsible for autoregressively generating tumor state predictions for future time points based on the context vector output by the encoder; the tumor evolution model uses a weighted combination of mean squared error loss function and KL divergence loss function as the optimization objective during the training phase.

5. The AI-assisted remote consultation system for testicular nucleocarcinoma according to claim 1, characterized in that, The deep reinforcement learning network employs an actor-critic architecture. The actor network is responsible for outputting diagnostic action strategies based on the current system state, where the diagnostic action strategy is a probability distribution across all possible diagnostic actions. The critic network is responsible for evaluating the value function of the current state, i.e., the long-term expected cumulative reward for executing the actor network's strategy. The reward signal of the deep reinforcement learning network consists of three parts: a consistency reward between the diagnostic action and the real expert's diagnostic record, a delayed reward for the diagnostic result being verified as correct by subsequent clinical follow-up, and a penalty for uncertainty in the diagnostic decision. The deep reinforcement learning network is trained using a proximal policy optimization algorithm.

6. The AI-assisted remote consultation system for testicular nucleocarcinoma according to claim 1, characterized in that, The version control mechanism operates as follows: the system assigns a globally unique identifier to each diagnostic session; each intermediate diagnostic conclusion or final decision generated within the session is considered a version node, and nodes are connected by directional edges to represent the derivation path of the diagnostic logic; each node stores the decision content, decision timestamp, a list of characteristic sources of the decision basis, and decision confidence level; when multiple experts consult on the same case, the system detects and highlights the divergence points between the decision paths of different experts, and automatically initiates an evidence-based conflict resolution process.

7. The AI-assisted remote consultation system for testicular nucleocarcinoma according to claim 1, characterized in that, It also includes a distributed model update and knowledge distillation subsystem; the distributed model update and knowledge distillation subsystem is deployed on a cloud server cluster and is responsible for periodically collecting desensitized diagnostic data and expert feedback from various access medical institutions; Using this new data, the core AI model in the system is incrementally updated through a federated learning framework; the knowledge distillation component compresses the complex policy network learned from the expert diagnostic logic deep simulation module into a lightweight student network, which can be deployed to resource-constrained edge computing devices.

8. The AI-assisted remote consultation system for testicular nucleocarcinoma according to claim 1, characterized in that, It also includes a security and privacy protection layer; the security and privacy protection layer is integrated into all data communication and storage links; all data leaving the user terminal is encrypted using encryption algorithms that comply with national cryptographic management standards before transmission; patient identification information stored in the system database is replaced with irreversible hash values; Sensitive medical data is divided into multiple data fragments and stored on different physical servers.

9. The AI-assisted remote consultation system for testicular nucleocarcinoma according to claim 2, characterized in that, In the outlier detection and correction sub-step, the system calculates an anomaly score for each data instance, which is based on the path length from the root node to the leaf node of the instance in the isolation tree; when the anomaly score exceeds a preset threshold of 0.95, the system automatically triggers the correction procedure.

10. The AI-assisted remote consultation system for testicular nucleocarcinoma according to claim 3, characterized in that, The attention weight calculation module is implemented by a three-layer perceptron. The input is the concatenated multimodal feature vector and its context metadata. The number of hidden layer neurons is 256 and 128 respectively. The output layer generates the attention weight values ​​corresponding to the four modalities through the Softmax function.

Images