Tumor risk prediction method and system for assisting prostate needle biopsy decision based on deep learning

Abstract

The invention discloses a tumor risk prediction method and system for assisting prostate needle biopsy decision based on deep learning, and belongs to the technical field of medical artificial intelligence. According to the system, a four-dimensional dynamic fusion and double closed-loop optimization architecture is innovatively constructed, and accurate prediction of prostate cancer risk and intelligent planning of a puncture path are realized through five core modules; a data acquisition module integrates preoperative MRI, intraoperative ultrasound (including SWE and blood flow data), driving gene data and puncture instrument sensing data; the intra-operative real-time multi-modal fusion module adopts a space-time attention Transform network and a CNN-GRU network to realize dynamic feature fusion; the image-genome association modeling module establishes association between a molecular mechanism and an image phenotype through a knowledge graph and a GAT network; the unsupervised field generalization module is combined with a PLDG mechanism and federated learning to improve the cross-center adaptability of the model; and the reinforcement learning real-time navigation module realizes dynamic adjustment of a puncture path based on a PPO algorithm.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of artificial intelligence, in particular to a tumor risk prediction method and system for assisting prostate biopsy decision based on deep learning. BACKGROUND

[0002] Prostate cancer is the second most common cancer and the fifth leading cause of cancer death in men worldwide, and has become a serious public health problem threatening men's health. Local prostate cancer is usually asymptomatic, and patients are difficult to detect by themselves. When obvious symptoms appear, the disease has often developed to an incurable stage, so early and accurate diagnosis is crucial for the treatment and prognosis of prostate cancer. There are still many problems to be solved in the current clinical commonly used prostate biopsy technique: during the traditional transrectal biopsy, patients suffer from severe pain due to the rich nerve endings in the rectal mucosa, and the body moves involuntarily, resulting in reduced accuracy of puncture positioning. In addition, a large number of bacteria in the rectum can cause infection through the puncture channel, and clinical data shows that the incidence of infection after transrectal puncture is significantly higher than that of other puncture methods. More importantly, due to the limitation of puncture angle and field of view, the traditional method has a high missed diagnosis rate of 22% for small lesions (≤5mm) in the special parts of the prostate, such as the anterior, peripheral and apex, which often leads to delayed early diagnosis of cancer.

[0003] In the aspect of image-guided technology, the existing multi-modal fusion method has obvious limitations. Although the fusion technology of multi-parameter magnetic resonance imaging (MP-MRI) and ultrasound has been used to assist puncture, it mainly relies on plane registration or simple rigid registration, and uses anatomical landmarks such as puborectalis muscle for matching, which cannot realize real-time dynamic adjustment during operation, and the registration accuracy is greatly affected by respiratory motion and tissue deformation. Although ultrasound elastography can judge the lesion properties through tissue hardness, traditional strain elastography can only reflect the instantaneous hardness of the tissue, and it is not combined with hemodynamic characteristics for analysis, so the diagnostic efficiency is limited. The existing risk prediction model also faces three technical bottlenecks: first, it does not integrate genomic data, which cannot explain the clinical phenomenon of "similar images but different prognoses", such as 15%-20% of patients with PI-RADS4 being actually low-risk cancer; second, the heterogeneity of multi-center data leads to poor generalization ability of the model, and the AUC value decreases by 12%-18% when tested across centers; third, the puncture path planning is a static pre-planning mode, which cannot respond to intraoperative tissue deformation, and when the puncture needle path deviates by more than 3mm, the target lesion may be missed. Therefore, it is an urgent need to develop an intelligent auxiliary system that can realize real-time fusion of multi-modal data, integrate molecular mechanisms, adapt to multi-center differences, and have dynamic navigation capability, to improve the accuracy and safety of prostate biopsy. SUMMARY

[0004] The application provides a tumor risk prediction method and system for assisting prostate biopsy decision based on deep learning, aiming at solving the problems in the background art.

[0005] The application is a tumor risk prediction system for assisting prostate biopsy decision based on deep learning, which is implemented as follows: A data acquisition module is used to acquire multi-source data related to prostate biopsy, including preoperative MRI image data, Intraoperative ultrasound data, including SWE data of ultrasound shear wave elastography, Doppler blood flow time series data, Prostate cancer driver gene data, including PTEN, TP53, and AR-V7 gene expression data, Puncture instrument sensing data, including tissue resistance data of a force sensor and needle tip coordinate data of a position sensor; An intraoperative real-time multi-modal fusion module is connected to the data acquisition module, and sub-pixel level registration of the preoperative MRI image data and intraoperative ultrasound data is realized through a spatio-temporal attention Transformer network, the Young's modulus value of the SWE data, the Doppler blood flow time series data, and the preoperative 12-24 hour pretreatment completed driver gene expression characteristics are integrated, the PTEN, TP53, and AR-V7 genes are detected based on preoperative peripheral blood microfluidic liquid biopsy, and the detection results are stored in the system database after standardization processing, the 'fast-in fast-out' malignant perfusion characteristics in the Doppler blood flow time series data are extracted through a low-delay interface call with a delay of ≤10 ms, a lightweight CNN-GRU network is used, the malignant perfusion characteristics, MRI static characteristics, and gene expression characteristics are adaptively weighted and fused, and the lesion risk heat map is dynamically updated; An image-genome association modeling module is connected to the data acquisition module, a knowledge graph of prostate cancer driver genes and imageomics features is constructed, a graph attention network is used to locate the image phenotype corresponding to the gene variation, a 'gene-image' joint embedding space is trained through contrast learning, and the gene expression amount is predicted based on the joint embedding space; An unsupervised domain generalization module is connected to the intraoperative real-time multi-modal fusion module and the image-genome association modeling module, respectively, a prompt-driven latent domain generalization mechanism is introduced, pseudo-domain labels are generated by k-means clustering of ViT network shallow style features, cross-center knowledge is shared through a domain prompt generator, a model parameter aggregation strategy under a federated learning architecture is used, and the influence of multi-center device parameter and operation differences on model performance is reduced; The reinforcement learning real-time navigation module is connected to the data acquisition module and the intraoperative real-time multimodal fusion module. It trains the puncture path decision model based on deep reinforcement learning, takes the tissue resistance data and needle tip coordinate data as the model state input, uses "target area hit accuracy + path length + bleeding risk" as the reward function, and adopts the ProximalPolicyOptimization algorithm. When the needle tip deviates from the planned path by more than a preset threshold, path replanning is triggered. The results output module is connected to the intraoperative real-time multimodal fusion module, the image-genome association modeling module, and the reinforcement learning real-time navigation module, respectively, and is used to output tumor risk prediction results, including the localization of small lesions and Gleason grade prediction; puncture path planning scheme and interpretable visualization results, including gene association heatmap and decision trajectory map.

[0006] Preferably, the Young's modulus value of the SWE data is set with a malignancy judgment threshold: Emean ≥ 47.005 kPa; the lightweight CNN-GRU network includes 3 CNN convolutional layers and 2 GRU recurrent layers, the CNN convolutional layers are used to extract blood flow spatial features, and the GRU recurrent layers are used to extract blood flow temporal features.

[0007] Preferably, in the image-genome association modeling module, the contrastive learning training process uses a triplet loss function, the dimension of the "gene-image" joint embedding space is 512, and the average absolute error of gene expression level prediction based on image features is ≤0.12.

[0008] Preferably, in the unsupervised domain generalization module, the model parameter aggregation strategy of the federated learning architecture adopts the weighted average method, with the weight coefficient = (center sample size / total sample size) × α + (center data quality score / average quality score) × (1-α), where α is the balance coefficient, the data quality score is based on the consistency of lesion annotation and the evaluation of equipment calibration level, and the cross-center model performance degradation is controlled to AUC decrease ≤ 5%.

[0009] Preferably, in the reinforcement learning real-time navigation module, the preset threshold is 2mm; the measurement accuracy of the force sensor is ≤0.01N, and the positioning accuracy of the position sensor is ≤0.1mm; the training iterations of the puncture path decision model are ≥10,000 rounds, and the path adjustment response time is <100ms; the bleeding risk score is calculated as (tissue resistance / 0.5N)×0.4 + (puncture depth / 20mm)×0.3 + (ultrasound vascular density score / 4 points)×0.3, where tissue resistance >0.5N, puncture depth >20mm, and vascular density score ≥3 points are high-risk thresholds.

[0010] Preferably, a tumor risk prediction method based on deep learning to assist in prostate biopsy decision-making includes the following steps: S1. Data Acquisition and Preprocessing: Acquire multi-source data related to prostate puncture, including preoperative MRI image data, intraoperative ultrasound data, driver gene data, and puncture instrument sensor data, and perform standardized preprocessing on the multi-source data; S2. Intraoperative real-time multimodal fusion: Subpixel-level registration of preoperative MRI image data and intraoperative ultrasound data is achieved through a spatiotemporal attention Transformer network. A lightweight CNN-GRU network is used to extract the "fast in and fast out" malignant perfusion features of Doppler blood flow time series data. The Young's modulus value of SWE data is integrated, and the malignant perfusion features are adaptively weighted and fused with MRI static features to dynamically update the lesion risk heat map. S3. Image-genome association modeling: Construct a knowledge graph of prostate cancer driver genes and radiomics features, use GAT network to locate the image phenotypes corresponding to gene variants, train the "gene-image" joint embedding space through comparative learning, and predict gene expression levels based on the joint embedding space; S4. Unsupervised Domain Generalization Optimization: k-means clustering is performed on the shallow style features of the ViT network to generate pseudo-domain labels. Domain hint generators using the PLDG mechanism share cross-center knowledge. The parameter aggregation strategy of federated learning is adopted to optimize the model and reduce the impact of heterogeneity of multi-center data. S5. Reinforcement Learning Real-time Navigation: Based on deep reinforcement learning, a puncture path decision model is trained. Tissue resistance data and needle tip coordinate data are input. The reward function is "target area hit accuracy + path length + bleeding risk". The PPO algorithm is used. When the needle tip deviates from the planned path by more than 2mm, path replanning is triggered. S6. Risk Prediction and Result Output: Integrate the lesion heat map from step S2, the gene expression prediction results from step S3, and the path planning scheme from step S5 to output the tumor risk level and interpretable visualization results.

[0011] Preferably, in step S2, the spatiotemporal attention Transformer network has 6 encoder layers and 8 attention heads, and the pixel deviation of the registered image is ≤0.5 pixels; the adaptive weight fusion adopts an attention mechanism, and the weight allocation range of static features and dynamic features is 0.3-0.7.

[0012] Preferably, in step S3, the knowledge graph contains the association between 1000+ radiomics features and 30+ gene variant sites; the training batch size of the comparative learning is 32, the learning rate is 1e-4, and the MAE of gene expression prediction after training is ≤0.12.

[0013] Preferably, in step S4, the number of clusters in the k-means clustering is k=5-8, the number of participating centers in the federated learning is ≥3, the parameter aggregation cycle is once every 100 rounds of training iteration, and the model AUC in cross-center testing is ≥0.91.

[0014] Due to the adoption of the above scheme, the beneficial effects of the present invention are: improving the detection rate of small lesions: by achieving sub-pixel-level image registration through the spatiotemporal attention Transformer, combined with the SWE elasticity feature and the "fast in and fast out" blood flow feature, the detection rate of ≤5mm small lesions is increased from 68% in the prior art to 89%, especially solving the problem of missed diagnosis of lesions in difficult-to-penetrate areas such as the apex and anterior part of the prostate. Enhancing the accuracy of risk prediction: Image-genome association modeling achieved a prediction accuracy of 92.7% for high-risk cancers (Gleason ≥ 8), an improvement of 15.4% compared to traditional imaging models. By explaining the phenomenon of "similar images but different prognoses" through gene association heatmaps, the unnecessary biopsy avoidance rate of PI-RADS3 grouping increased from 28.8% to 41.3%. Enhancing the robustness of cross-center models: A domain generalization scheme combining PLDG mechanism and federated learning was adopted, achieving an average AUC of 0.91 on test sets of three hospitals of different levels. This reduces the AUC decay by more than 70% compared to existing technologies, and solves the problem of model performance instability caused by differences in equipment parameters and operations in multiple centers. Optimizing the safety and efficiency of puncture navigation: The reinforcement learning real-time navigation system reduces the puncture path adjustment response time to <100ms, shortening the operation time to 25 minutes (compared to an average of 31 minutes with existing technologies). Dynamic bleeding risk assessment reduces the incidence of complications (such as bleeding and infection) to <3%, a significant improvement over traditional transrectal puncture. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] A deep learning-based tumor risk prediction system for assisting prostate biopsy decision-making includes: The data acquisition module is used to acquire multi-source data related to prostate biopsy, including preoperative MRI image data, intraoperative ultrasound data (including ultrasound shear wave elastography (SWE) data and Doppler blood flow time-series data), prostate cancer driver gene data (including PTEN, TP53, and AR-V7 gene expression data), and biopsy instrument sensing data (including tissue resistance data from force sensors and needle tip coordinate data from position sensors). The intraoperative real-time multimodal fusion module is connected to the data acquisition module. It achieves sub-pixel-level registration of the preoperative MRI image data and intraoperative ultrasound data through a spatiotemporal attention Transformer network. It integrates the Young's modulus value of the SWE data, Doppler blood flow time series data, and driver gene expression features preprocessed 12-24 hours before surgery. It detects PTEN, TP53, and AR-V7 genes based on preoperative peripheral blood microfluidic biopsy. The detection results are stored in the system database after standardization. During surgery, it is called through a low-latency interface with a latency of ≤10ms. A lightweight CNN-GRU network is used to extract the 'fast in and fast out' malignant perfusion features in the Doppler blood flow time series data. The malignant perfusion features, MRI static features, and gene expression features are adaptively weighted and fused to dynamically update the lesion risk heat map. The image-genome association modeling module is connected to the data acquisition module to construct a knowledge graph of prostate cancer driver genes and radiomics features. It uses graph attention network (GAT) to locate the image phenotypes corresponding to gene variations, trains the "gene-image" joint embedding space through comparative learning, and realizes the prediction of gene expression levels from image features based on the joint embedding space. The unsupervised domain generalization module is connected to the intraoperative real-time multimodal fusion module and the image-genome association modeling module, respectively. It introduces a prompt-driven latent domain generalization (PLDG) mechanism to generate pseudo-domain labels by k-means clustering of shallow style features of the ViT network. It shares cross-center knowledge through a domain prompt generator and adopts a model parameter aggregation strategy under a federated learning architecture to reduce the impact of differences in multi-center equipment parameters and operations on model performance. The reinforcement learning real-time navigation module is connected to the data acquisition module and the intraoperative real-time multimodal fusion module. Based on deep reinforcement learning, the puncture path decision model is trained. The tissue resistance data and needle tip coordinate data are used as the model state input. The reward function is "target area hit accuracy + path length + bleeding risk". The Proximal Policy Optimization (PPO) algorithm is used. When the needle tip deviates from the planned path by more than a preset threshold, path replanning is triggered. The results output module is connected to the intraoperative real-time multimodal fusion module, the image-genome association modeling module, and the reinforcement learning real-time navigation module, respectively, and is used to output tumor risk prediction results (including small lesion localization and Gleason grade prediction), puncture path planning scheme, and interpretable visualization results (including gene association heatmap and decision trajectory map). In the intraoperative real-time multimodal fusion module, the Young's modulus value of the SWE data is set with a malignancy judgment threshold: Emean ≥ 47.005 kPa; the lightweight CNN-GRU network includes 3 CNN convolutional layers (convolutional kernel sizes are 3×3, 3×3, and 5×5 respectively) and 2 GRU recurrent layers. The CNN convolutional layers are used to extract blood flow spatial features, and the GRU recurrent layers are used to extract blood flow temporal features. In the image-genome association modeling module, the contrastive learning training process uses a triplet loss function, the dimension of the "gene-image" joint embedding space is 512-dimensional, and the mean absolute error (MAE) of gene expression level prediction based on image features is ≤0.12. In the unsupervised domain generalization module, the model parameter aggregation strategy of the federated learning architecture adopts the weighted average method, with the weight coefficient = (center sample size / total sample size) × α + (center data quality score / average quality score) × (1-α), where α is the balance coefficient, the data quality score is based on the consistency of lesion annotation and the evaluation of equipment calibration level, and the cross-center model performance degradation is controlled to AUC decrease ≤ 5%. In the reinforcement learning real-time navigation module, the preset threshold is 2mm; the measurement accuracy of the force sensor is ≤0.01N, and the positioning accuracy of the position sensor is ≤0.1mm; the training iterations of the puncture path decision model are ≥10,000 rounds, and the path adjustment response time is <100ms; the bleeding risk score is calculated as (tissue resistance / 0.5N)×0.4+(puncture depth / 20mm)×0.3+(ultrasound vascular density score / 4 points)×0.3, where tissue resistance >0.5N, puncture depth >20mm, and vascular density score ≥3 points are high-risk thresholds. A deep learning-based method for predicting tumor risk in prostate biopsy decisions includes the following steps: S1. Data Acquisition and Preprocessing: Acquire multi-source data related to prostate biopsy, including preoperative MRI image data, intraoperative ultrasound data (including SWE data and Doppler blood flow time-series data), driver gene data, and biopsy instrument sensor data. Perform standardized preprocessing on the multi-source data (MRI images are compressed to 224×224 pixels, and the sampling frequency of Doppler blood flow time-series data is uniformly set to 10Hz). S2. Intraoperative real-time multimodal fusion: Subpixel-level registration of preoperative MRI image data and intraoperative ultrasound data is achieved through a spatiotemporal attention Transformer network. A lightweight CNN-GRU network is used to extract the "fast in and fast out" malignant perfusion features of Doppler blood flow time series data. The Young's modulus value (Emean ≥ 47.005 kPa) of SWE data is integrated. The malignant perfusion features are adaptively weighted and fused with MRI static features to dynamically update the lesion risk heat map. S3. Image-genome association modeling: Construct a knowledge graph of prostate cancer driver genes (PTEN, TP53, AR-V7) and radiomics features, use GAT network to locate the image phenotypes corresponding to gene variants, train the "gene-image" joint embedding space through comparative learning, and predict gene expression levels based on the joint embedding space; S4. Unsupervised Domain Generalization Optimization: k-means clustering is performed on the shallow style features of the ViT network to generate pseudo-domain labels. Domain hint generators using the PLDG mechanism share cross-center knowledge. The parameter aggregation strategy of federated learning is adopted to optimize the model and reduce the impact of heterogeneity of multi-center data. S5. Reinforcement Learning Real-time Navigation: Based on deep reinforcement learning, a puncture path decision model is trained. Tissue resistance data and needle tip coordinate data are input. The reward function is "target area hit accuracy + path length + bleeding risk". The PPO algorithm is used. When the needle tip deviates from the planned path by more than 2mm, path replanning is triggered. S6. Risk Prediction and Result Output: Integrate the lesion heatmap from step S2, the gene expression prediction results from step S3, and the path planning scheme from step S5 to output the tumor risk level (including the location of small lesions <5mm and Gleason grade prediction) and interpretable visualization results (gene association heatmap and decision trajectory map). In step S2, the spatiotemporal attention Transformer network has 6 encoder layers and 8 attention heads, and the pixel deviation of the registered image is ≤0.5 pixels; the adaptive weight fusion adopts an attention mechanism, and the weight allocation range of static features and dynamic features is 0.3-0.7. In step S3, the knowledge graph contains the association between 1000+ radiomics features (including texture features, shape features, and grayscale features) and 30+ gene mutation sites; the training batch size of the contrastive learning is 32, the learning rate is 1e-4, and the MAE of gene expression prediction after training is ≤0.12. In step S4, the number of clusters in the k-means clustering is k=5-8, the number of participating centers in the federated learning is ≥3, the parameter aggregation period is once every 100 rounds of training iteration, and the model AUC in cross-center testing is ≥0.91. In step S5, the reward function formula of the puncture path decision model is: Reward = 0.6 × target hit accuracy + 0.2 × (1 - path length / maximum path threshold) + 0.2 × (1 - bleeding risk score), where the bleeding risk score is calculated based on tissue resistance data and puncture depth; the execution time of the path replanning is <100ms, and the target hit accuracy after replanning is ≥95%.

[0017] In this embodiment, the system includes a data acquisition module, an intraoperative real-time multimodal fusion module, an image-genome association modeling module, an unsupervised domain generalization module, a reinforcement learning real-time navigation module, and a result output module. These modules work together to achieve accurate prediction of prostate cancer risk and intelligent planning of puncture paths.

[0018] The data acquisition module is responsible for acquiring preoperative MRI image data, intraoperative ultrasound data (including ultrasound shear wave elastography (SWE) data and Doppler blood flow time-series data), prostate cancer driver gene data (including PTEN, TP53, and AR-V7 gene expression data), and puncture instrument sensor data (including tissue resistance data from force sensors and needle tip coordinate data from position sensors). The intraoperative real-time multimodal fusion module achieves sub-pixel-level registration of preoperative MRI and intraoperative ultrasound through a spatiotemporal attention Transformer network. It uses a lightweight CNN-GRU network to extract blood flow temporal features and performs adaptive weight fusion with SWE Young's modulus value and MRI static features to dynamically update the lesion risk heat map. The image-genome association modeling module constructs a knowledge graph of driving genes and radiomics features, uses graph attention network to locate the image phenotypes corresponding to gene variations, and trains the "gene-image" joint embedding space through contrastive learning to achieve gene expression prediction based on image features. The unsupervised domain generalization module employs a prompt-driven latent domain generalization mechanism and a federated learning architecture to reduce the impact of differences in multi-center device parameters and operations on model performance. The reinforcement learning real-time navigation module is based on a deep reinforcement learning training puncture path decision model, and combines force sensor and position sensor data to achieve real-time path adjustment. The results output module integrates the results from each module and outputs tumor risk prediction results, puncture path planning schemes, and interpretable visualization results. The method of the present invention includes the following steps: Data acquisition and preprocessing: Acquire multi-source data and perform standardization processing, including MRI images compressed to 224×224 pixels and Doppler blood flow time series data sampling frequency uniformly set to 10Hz; Intraoperative real-time multimodal fusion: Image registration is achieved through spatiotemporal attention Transformer, blood perfusion features are extracted and adaptively fused with other features; Image-genome association modeling: Construct a knowledge graph, use the GAT network to locate the image phenotype of gene variation, and train the joint embedding space to achieve gene expression prediction; Unsupervised domain generalization optimization: Model optimization through PLDG mechanism and federated learning parameter aggregation; Reinforcement learning real-time navigation: Train a path decision model and dynamically adjust the puncture path based on sensor data; Risk prediction and results output: Integrates results from multiple modules to output risk levels and visualization results. Specific Implementation Example 1: Data Acquisition and Preprocessing Equipment configuration: Preoperative multi-parameter MRI data were acquired using a Siemens Magnetom Skyra 3.0-T MRI scanner, including T2-weighted imaging, diffusion-weighted imaging, and dynamic contrast-enhanced sequences, with a slice thickness of 3 mm and no gaps in the scan. Intraoperative ultrasound data were acquired using a LOGIQE9 color Doppler ultrasound system equipped with an IC5-9-D cavity transducer, simultaneously acquiring B-mode ultrasound, shear wave elastography (SWE), and color Doppler blood flow data. The SWE sampling frequency was 5 Hz, and the blood flow data sampling frequency was 10 Hz. Gene Data Collection: Preoperative Gene Testing Procedure: 12-24 hours before surgery, 5 mL of peripheral blood is collected from the patient. Circulating tumor DNA (ctDNA) is separated using microfluidic liquid biopsy technology. Thermo Fisher Quant Studio 7 Flex real-time qPCR instrument is used to detect PTEN exon 5, TP53 exon 7, and AR-V7 splicing sites. A mutation abundance of ≥5% is considered positive. The test results are standardized using FPKM (fiber-per-kbps transcripts per million mapping reads) and then stored in the system's local database. Intraoperative correlation and retrieval: When intraoperative ultrasound locates a specific area of ​​the prostate, the system automatically matches the preoperative MRI partition label corresponding to that area, and retrieves the gene expression characteristics associated with that partition from the database based on the label, so as to achieve accurate matching between gene characteristics and real-time intraoperative image characteristics. The actual retrieval delay is ≤8ms. To prevent contamination and ensure data validity, EDTA anticoagulant tubes were used before peripheral blood collection to avoid hemolysis interference. Negative and positive controls (with known mutant plasmids) were set up during the testing process to ensure that the detection specificity was ≥95%, and the data validity rate reached 100% in 100 clinical trials.

[0019] Sensor configuration: The puncture instrument is equipped with a force sensor with an accuracy of ≤0.01N and an electromagnetic position sensor with a positioning accuracy of ≤0.1mm, with a sampling frequency of 100Hz. Preprocessing steps: MRI images were subjected to N4 bias correction to remove magnetic field inhomogeneities, rigidly registered to a standard template space, and then the region of interest was cropped and compressed to 224×224 pixels. Young's modulus mean (Emean) was extracted from SWE data, and Doppler blood flow data were denoised using adaptive median filtering. Time-intensity curves were extracted using wavelet transform. Example 2: Implementation Details of the Intraoperative Real-Time Multimodal Fusion Module Spatiotemporal attention Transformer registration: The network consists of 6 encoder layers, each with 8 attention heads. The input consists of MRI static features (512-dimensional) and ultrasound dynamic features (256-dimensional). A self-attention mechanism is used to learn intramodal spatiotemporal correlations, while a cross-attention mechanism captures intermodal correlations. After registration, the image pixel deviation is ≤0.5 pixels. The loss function combines normalized mutual information (NMI) with gradient consistency loss. The lightweight CNN-GRU network consists of three CNN convolutional layers (kernel sizes of 3×3, 3×3, and 5×5, with 64, 128, and 256 output channels, respectively) and two GRU recurrent layers (128 hidden units). The CNN layers extract spatial features of blood flow, while the GRU layers capture the temporal features of malignant perfusion with rapid inflow and outflow, ultimately outputting a 128-dimensional feature vector. Adaptive weight fusion: An attention mechanism is used to dynamically allocate weights, with the weights for static MRI features and dynamic blood flow features ranging from 0.3 to 0.7. When Emean ≥ 47.005 kPa in SWE data, the weight of elasticity features is automatically increased to enhance the ability to identify malignant lesions. Specifically, the update frequency is synchronized with the intraoperative ultrasound frame rate (10Hz, i.e., once every 100ms). Triggering conditions: When the change in Young's modulus value of SWE data exceeds 5 kPa, or the change in Doppler blood flow characteristics (such as peak flow velocity) exceeds 20%, an emergency update of the heat map is triggered (response time ≤ 50 ms).

[0020] Example 3: Image-genome association modeling Knowledge graph construction: It includes the association between 1000+ radiomics features (texture features, shape features, grayscale features) and 30+ gene variant sites. Initial associations are established through literature mining and clinical expert annotation, and knowledge representation learning is carried out using the TransE algorithm. GAT network structure: The input is an image feature matrix (n×512) and a gene feature matrix (m×200). The importance weights of nodes are learned through two graph attention layers. Each layer has 256 hidden units, 4 attention heads, and a dropout rate of 0.3. Contrastive learning training: A triplet loss function was used, with a batch size of 32, a learning rate of 1e-4, and 200 training epochs. The joint "gene-image" embedding space had a dimension of 512, which was mapped to a 256-dimensional space using a projection head to calculate similarity. After training, the mean absolute error (MAE) of gene expression prediction was ≤0.12. Example 4: Unsupervised Domain Generalization Module PLDG mechanism: k-means clustering (k=5-8) is performed on the style features of the first 4 layers of the ViT-Base network to generate pseudo-domain labels. The domain hint generator is a 3-layer MLP network, which generates 128-dimensional domain hint vectors and injects them into the attention module of each layer of the Transformer. Federated learning architecture: Employing a parameter server model, with participation from at least 3 medical centers, 50 local training rounds, and parameter aggregation occurring every 100 rounds. The aggregation strategy uses a weighted average method, with weight coefficients positively correlated with the data volume proportion of each center, and cross-center model performance degradation is controlled to an AUC decrease of ≤5%. Specifically, consistency in lesion labeling (Kappa value ≥ 0.8 is the full score of 100 points, and 20 points are deducted for each 0.1 lower value); - Equipment calibration level (100 points for annual calibration, 80 points for calibration within 6 months of expiration, and 50 points for calibration more than 6 months of expiration). Weighting calculation example: If α=0.6, center A has a sample size of 30% and a quality score of 90, center B has a sample size of 70% and a quality score of 80, and the average quality score is 85, then the weight of A = 30%×0.6+(90 / 85)×0.4≈0.35, and the weight of B = 70%×0.6+(80 / 85)×0.4≈0.65.

[0021] Example 5: Reinforcement Learning Real-time Navigation Module PPO algorithm configuration: The state space includes needle tip coordinates (x, y, z), tissue resistance, and target lesion coordinates; the action space includes the puncture angle adjustment (-5° to +5°) and needle depth (0-5mm). Discount factor γ=0.99, gae parameter λ=0.95, and strategy update clipping parameter ε=0.2. Reward function calculation: Reward = 0.6 × target accuracy + 0.2 × (1 - path length / maximum path threshold) + 0.2 × (1 - bleeding risk score). The bleeding risk score is calculated based on tissue resistance (>0.5N indicates high risk) and puncture depth (>20mm, risk increases). Dynamic replanning mechanism: Replanning is triggered when the needle tip deviates from the planned path by more than 2mm. A greedy algorithm combined with local trajectory optimization is used, and the path adjustment response time is less than 100ms. After replanning, the target area hit accuracy is ≥95%. The model training iterations are ≥10,000 rounds, and multi-environment parallel sampling is used to accelerate training.

[0022] The above description of the embodiments is intended to enable those skilled in the art to understand and use the present invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments. Improvements and modifications made by those skilled in the art based on the principles of the present invention without departing from the scope of the invention should be within the protection scope of the present invention. The above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based tumor risk prediction system for assisting prostate biopsy decision-making, characterized by, Comprise: a data acquisition module for acquiring prostate biopsy-related multi-source data, the multi-source data including preoperative MRI image data, intraoperative ultrasound data including ultrasound shear wave elastography (SWE) data, Doppler blood flow time series data, prostate cancer driver gene data including PTEN, TP53, AR-V7 gene expression data, puncture instrument sensing data including tissue resistance data of a force sensor and needle tip coordinate data of a position sensor; an intraoperative real-time multi-modal fusion module connected with the data acquisition module and configured to achieve sub-pixel level registration of the preoperative MRI image data and the intraoperative ultrasound data through a space-time attention Transformer network, integrate the Young's modulus value of the SWE data, the Doppler blood flow time series data, and the preoperative 12-24 hour pretreatment completed driver gene expression characteristics, which are based on preoperative peripheral blood microfluidic liquid biopsy detection of PTEN, TP53, and AR-V7 genes, and the detection results are stored in a system database after standardization processing, intraoperative call through a low-delay interface with a delay of ≤10 ms, extraction of the 'fast-in fast-out' malignant perfusion characteristics in the Doppler blood flow time series data using a lightweight CNN-GRU network, and adaptive weight fusion of the malignant perfusion characteristics, MRI static characteristics, and gene expression characteristics to dynamically update the lesion risk heat map; an image-genome association modeling module connected with the data acquisition module, which constructs a knowledge graph of prostate cancer driver genes and imageomics features, uses a graph attention network to locate the corresponding image phenotype of gene variation, trains a "gene-image" joint embedding space through contrastive learning, and realizes prediction of gene expression from image features based on the joint embedding space; an unsupervised domain generalization module connected with the intraoperative real-time multi-modal fusion module and the image-genome association modeling module, which introduces a prompt-driven latent domain generalization mechanism, performs k-means clustering on the shallow style features of the ViT network to generate pseudo-domain labels, shares cross-center knowledge through a domain prompt generator, and uses a model parameter aggregation strategy under a federated learning architecture to reduce the influence of multi-center device parameter and operation differences on model performance; a reinforcement learning real-time navigation module connected with the data acquisition module and the intraoperative real-time multi-modal fusion module, which trains a puncture path decision model based on deep reinforcement learning, uses the tissue resistance data and needle tip coordinate data as model state input, uses "target area hit accuracy + path length + bleeding risk" as a reward function, and uses the Proximal Policy Optimization algorithm to trigger path re-planning when the needle tip deviates from the planned path by more than a preset threshold; a result output module connected with the intraoperative real-time multi-modal fusion module, the image-genome association modeling module, and the reinforcement learning real-time navigation module, which outputs tumor risk prediction results including micro-lesion localization and Gleason grade prediction, puncture path planning scheme, and interpretable visualization results including gene association heat map and decision trajectory graph.

2. The system of claim 1, wherein, In the intraoperative real-time multi-modal fusion module, the Young's modulus value of the SWE data is set with a malignant judgment threshold value: Emean≥47.005kPa; the light-weight CNN-GRU network comprises 3 layers of CNN convolution layers and 2 layers of GRU recurrent layers, the CNN convolution layer is used to extract blood flow spatial features, and the GRU recurrent layer is used to extract blood flow time sequence features.

3. The system of claim 1, wherein, In the image-genome correlation modeling module, the contrast learning training process adopts a triplet loss function, the dimension of the "gene-image" joint embedding space is 512, and the average absolute error of predicting gene expression based on image features is ≤0.

12.

4. The system of claim 1, wherein, In the unsupervised domain generalization module, the model parameter aggregation strategy of the federated learning architecture adopts a weighted average method, the weight coefficient = (center sample amount / total sample amount) × α + (center data quality score / average quality score) × (1-α), wherein α is a balance coefficient, the data quality score is evaluated based on lesion annotation consistency and equipment calibration level, and the cross-center model performance decay control is that the AUC drop amplitude is ≤5%.

5. The system of claim 1, wherein, In the reinforcement learning real-time navigation module, the preset threshold is 2mm; the measurement accuracy of the force sensor is ≤0.01N, and the positioning accuracy of the position sensor is ≤0.1mm; the training iteration number of the puncture path decision model is ≥10,000 rounds, and the path adjustment response time is <100ms; the bleeding risk score = (tissue resistance / 0.5N) × 0.4 + (puncture depth / 20mm) × 0.3 + (ultrasound blood vessel density score / 4 points) × 0.3, wherein tissue resistance>0.5N, puncture depth>20mm, and blood vessel density score≥3 points are high-risk threshold values.

6. The method of claim 1-5, wherein, The method comprises the following steps: S1. Data acquisition and preprocessing: acquiring multi-source data related to prostate puncture, the multi-source data comprising preoperative MRI image data, intraoperative ultrasound data, driver gene data, and puncture instrument sensing data, and performing standardization preprocessing on the multi-source data; S2. Intraoperative real-time multi-modal fusion: realizing sub-pixel level registration of preoperative MRI image data and intraoperative ultrasound data through a spatio-temporal attention Transformer network, extracting "fast-in fast-out" malignant perfusion features of Doppler blood flow time sequence data through a light-weight CNN-GRU network, integrating Young's modulus value of SWE data, adaptively fusing the malignant perfusion features and MRI static features, and dynamically updating a lesion risk heat map; S3. Image-genome correlation modeling: constructing a knowledge graph of prostate cancer driver genes and imageomics features, positioning the corresponding image phenotypes of gene variations through a GAT network, training a "gene-image" joint embedding space through contrast learning, and predicting gene expression based on the joint embedding space; S4. Unsupervised domain generalization optimization: performing k-means clustering on the shallow style features of the ViT network to generate pseudo domain labels, sharing cross-center knowledge through a domain prompt generator of a PLDG mechanism, optimizing the model through a parameter aggregation strategy of federated learning, and reducing the influence of multi-center data heterogeneity; S5. Reinforcement learning real-time navigation: based on deep reinforcement learning to train the puncture path decision model, input the tissue resistance data and needle tip coordinate data, take the "target area hit accuracy + path length + bleeding risk" as the reward function, use PPO algorithm, and trigger path re-planning when the needle tip deviates from the planned path > 2mm; S6. Risk prediction and result output: fuse the lesion heat map in step S2, the gene expression prediction result in step S3, and the path planning scheme in step S5, and output the tumor risk level and the interpretable visualization result.

7. The method of claim 6, wherein, In step S2, the number of encoder layers of the spatio-temporal attention Transformer network is 6, the number of attention heads is 8, and the pixel deviation of the registered image is ≤0.5 pixels; the adaptive weight fusion adopts the attention mechanism, and the weight distribution range of the static feature and the dynamic feature is 0.3-0.

7.

8. The method of claim 6, wherein, In step S3, the knowledge graph contains the association relationship of 1000+ imageomics features and 30+ gene mutation sites; the training batch size of the contrast learning is 32, the learning rate is 1e-4, and the MAE of the gene expression prediction after training is ≤0.

12.

9. The method of claim 6, wherein, In step S4, the number of clusters k of the k-means clustering is 5-8, the number of participating centers of the federated learning is ≥3, the parameter aggregation period is every 100 training iterations, and the model AUC of the cross-center test is ≥0.91.