A method and system for ultrasound video diagnosis of the prostate

Abstract

The application discloses a prostate ultrasonic video diagnosis method and system, wherein each frame of a prostate ultrasonic video is pretreated, multi-scale spatial morphological features in the pretreated prostate ultrasonic video are captured, a higher weight is given to the region, uncertainty perception is enhanced, the multi-scale spatial features are mapped back to the original image space, a prostate gland pixel-level segmentation mask is generated, the prostate gland pixel-level segmentation mask and the pretreated prostate ultrasonic video are input into a double-path attention fusion model, information in the spatial and temporal dimensions is extracted in parallel and is deeply cooperated and fused, finally, a key frame sequence is recognized, a comprehensive depth feature vector is generated through fusion, the significance of prostate cancer is finally judged, and a final diagnosis score is output.

Description

Technical Field

[0001] This invention belongs to the field of ultrasound data analysis technology, specifically relating to a method and system for prostate ultrasound video diagnosis. Background Technology

[0002] Prostate cancer (PCa) poses a significant threat to men's health worldwide, and early and accurate diagnosis is crucial for improving patient prognosis. Traditional transrectal ultrasound (TRUS)-guided prostate biopsy is currently the "gold standard" for clinical diagnosis of prostate cancer, but this method has significant limitations.

[0003] In recent years, the application of AI in the field of prostate cancer ultrasound diagnosis has evolved from early image preprocessing and automatic segmentation to more complex tasks such as cancer classification and grading, demonstrating enormous potential. Currently, the mainstream AI-assisted diagnostic techniques primarily focus on analysis using multi-parametric magnetic resonance imaging (mpMRI). Although mpMRI-based AI technology has made groundbreaking progress, its high equipment costs and examination fees limit its widespread adoption in primary healthcare institutions.

[0004] In contrast, ultrasound technology plays an indispensable role in the diagnosis and treatment of prostate diseases due to its affordability, convenience, and real-time capabilities. Therefore, applying AI to ultrasound image analysis, especially utilizing real-time acquired ultrasound video, has become an important research direction for improving the accessibility of diagnosis.

[0005] Existing ultrasound-based AI-assisted diagnostic technologies also face common limitations. First, their performance heavily relies on high-quality labeled data, but the labeling process for medical image data is extremely time-consuming and costly, which to some extent restricts model training and optimization. Second, many existing systems still suffer from a "black box" problem in their design; the model's decision-making process lacks transparency, making it difficult to explain the diagnostic basis to clinicians. This is one of the biggest obstacles to the clinical translation of AI technology in the medical field. Finally, most existing technologies rely on the analysis of static images, failing to fully utilize the rich dynamic information contained in ultrasound videos.

[0006] In summary, the following technical problems exist:

[0007] When processing prostate ultrasound videos, there is a general reliance on the analysis of static images, failing to fully utilize the dynamic information contained in the video. This process often requires manual intervention, such as manually segmenting the gland outline or identifying keyframes. This is not only time-consuming and labor-intensive but also introduces human error, seriously affecting the consistency and efficiency of diagnosis.

[0008] The diagnosis of prostate cancer is multidimensional, encompassing both the static spatial structure of the prostate gland and its morphological changes and blood flow characteristics during dynamic examination. However, most existing AI models employ a single feature extraction path, making it difficult to collaboratively model these complex spatiotemporal dependencies. For example, while some models incorporate attention mechanisms, they still face challenges in interpreting their complex decision-making processes.

[0009] In the medical field, even a high-performing "black box" model may be rejected by doctors because they cannot understand its decision-making logic. When receiving AI-assisted diagnosis, clinicians not only need an accurate result, but also a clear and verifiable explanation to confirm whether the model's decision-making basis is consistent with their own clinical judgment. Summary of the Invention

[0010] The purpose of this invention is to systematically overcome the three core technical challenges currently faced by prostate ultrasound video analysis technology in clinical applications, as mentioned in the background section, and to resolve the shortcomings such as these technical problems restricting the performance of AI models and becoming key obstacles to their transition from laboratory to clinical practice.

[0011] To achieve the above-mentioned objectives, the present invention provides a method for prostate ultrasound video diagnosis, comprising the following steps: Image compensation was performed on each frame of the prostate ultrasound video for inter-frame motion caused by heartbeat, respiration and probe movement, and temporal smoothing filtering was performed. The quality of the processed prostate ultrasound video was then evaluated to obtain the preprocessed prostate ultrasound video. By utilizing a Transformer-based encoder-decoder structure and introducing a weighted loss function for difficult regions, multi-scale spatial morphological features in the preprocessed prostate ultrasound video are captured and given higher weights to enhance uncertainty perception. The multi-scale spatial features are mapped back to the original image space to generate a pixel-level segmentation mask for the prostate gland. The pixel-level segmentation mask for the prostate gland and the preprocessed prostate ultrasound video are input into a dual-path attention fusion model to extract information in both spatial and temporal dimensions in parallel and perform deep collaboration and fusion. Finally, keyframe sequences are identified and fused to generate a comprehensive deep feature vector that represents the prostate state and contains both spatial and temporal information. The significance of prostate cancer is finally determined based on comprehensive deep feature vectors, and a final diagnostic score is output.

[0012] Preferably, the step of inputting the pixel-level segmentation mask of the prostate gland and the preprocessed prostate ultrasound video into a dual-path attention fusion model, and extracting information in the spatial and temporal dimensions in parallel and performing deep collaboration and fusion, includes the following steps: The pixel-level segmentation mask of the prostate gland is input into the spatial path. With internal attention to enhance the structural relationship between different regions of the gland, the static spatial structure features inside the gland are extracted in detail. The preprocessed prostate ultrasound video is input into the time path, and the dynamic temporal evolution characteristics of the prostate gland in the video are modeled by capturing the dynamic correlation between different frames in the video sequence through internal attention. If the spatial path identifies a suspicious structure in a specific region, then cross-attention guides the temporal path to focus on analyzing the dynamic changes of that region at different time points in the video sequence. If the time path identifies a dynamic anomaly at a specific time, then cross-attention guides the spatial path to focus on identifying and extracting static spatial structure features. Finally, keyframe sequences were identified and fused to generate a comprehensive deep feature vector that represents the prostate state and contains both spatial and temporal information.

[0013] The present invention also provides a prostate ultrasound video diagnostic system, employing a prostate ultrasound video diagnostic method as described above, the system comprising: The data access and preprocessing module is used to perform artifact denoising, brightness and contrast improvement on each frame of the prostate ultrasound video, perform image compensation for inter-frame motion caused by heartbeat, breathing and probe movement, and perform temporal smoothing filtering. It also performs quality assessment on the processed prostate ultrasound video, identifies and removes low-quality frames caused by equipment failure, image blur or improper operation, and obtains the preprocessed prostate ultrasound video. The intelligent processing module utilizes a Transformer-based encoder-decoder structure, introduces a weighted loss function for difficult regions, captures multi-scale spatial morphological features in the preprocessed prostate ultrasound video, assigns them higher weights to enhance uncertainty perception, maps the multi-scale spatial features back to the original image space, generates a pixel-level segmentation mask for the prostate gland, and inputs the pixel-level segmentation mask and the preprocessed prostate ultrasound video into a dual-path attention fusion model. It extracts information in the spatial and temporal dimensions in parallel and performs deep collaboration and fusion, finally identifying key frame sequences and fusing them to generate a comprehensive deep feature vector that represents the prostate state and contains both spatial and temporal information. The discrimination decision module is used to make a final judgment on the significance of prostate cancer based on a comprehensive deep feature vector and output the final diagnostic score.

[0014] This invention provides a method and system for prostate ultrasound video diagnosis. By preprocessing each frame of the prostate ultrasound video, multi-scale spatial morphological features in the preprocessed video are captured, and these features are assigned higher weights to enhance uncertainty perception. The multi-scale spatial features are mapped back to the original image space to generate a pixel-level segmentation mask for the prostate gland. The pixel-level segmentation mask and the preprocessed prostate ultrasound video are input into a dual-path attention fusion model, which extracts information in both spatial and temporal dimensions in parallel and performs deep collaboration and fusion. Finally, keyframe sequences are identified, and a comprehensive deep feature vector is generated. The saliency of prostate cancer is then determined, and a final diagnostic score is output. Detailed Implementation

[0015] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0016] An embodiment of the present invention provides a method for prostate ultrasound video diagnosis, comprising the following steps: Using the ECC image alignment algorithm, each frame of the video is denoised to eliminate artifacts introduced by equipment, operation, or patient movement. Histogram equalization and other techniques are used to standardize and enhance the images to ensure that the images in the input model are consistent in terms of brightness, contrast, etc. Based on the preset algorithm and clinical expert experience, the video sequence is quality assessed, and low-quality frames caused by equipment failure, image blurring, or improper operation are automatically identified and removed, retaining only clear and complete video segments that can be analyzed.

[0017] The ECC image alignment algorithm process is as follows: (1) Preprocessing and image enhancement: Each frame of the input ultrasound video is preprocessed, the BGR color image is converted into a grayscale image, and histogram equalization is applied to improve the image contrast.

[0018] (2) Motion compensation and image alignment: For the inter-frame motion caused by heartbeat, breathing and probe movement in ultrasound video, an image registration algorithm based on enhanced correlation coefficient (ECC) is used for compensation.

[0019] (3) Temporal smoothing filter: In order to suppress random noise and enhance stable anatomical structures, an exponential moving average (EMA) filter is used for temporal smoothing.

[0020] Utilizing an advanced Transformer-based encoder-decoder architecture, precise prostate gland segmentation is performed on each frame of the video, generating a high-quality segmentation mask. This enables prostate gland segmentation based on uncertainty perception, not only to locate the anatomical boundaries of the prostate but also to extract its spatial morphological features, providing precise anatomical context for subsequent analysis. After gland segmentation, dynamic analysis is performed on the entire video. By designing specialized attention mechanisms or feature clustering algorithms, keyframe sequences containing key prostate anatomical structures, abnormal morphological changes, or important hemodynamic information are intelligently identified. This step aims to extract the most diagnostically valuable information from the lengthy video, significantly improving the efficiency of subsequent analysis. Two parallel feature extraction paths are designed: one focuses on analyzing the static spatial structural features of the prostate gland, while the other aims to capture the dynamic temporal evolution features of the gland throughout the video. The two paths work in tandem through a dual-path (internal and cross) attention mechanism, ultimately fusing to generate a comprehensive deep feature vector that fully characterizes the prostate state and contains both spatial and temporal information.

[0021] This invention utilizes advanced encoder-decoder structures such as SegFormer for prostate gland segmentation. The encoder employs a hybrid convolutional and Transformer architecture to efficiently capture multi-scale spatial features in the image, while the decoder accurately maps these abstract features back to the original image space, generating pixel-level segmentation masks for the glands. However, the key innovation of this invention lies not in the simple reuse of the basic architecture, but in the introduction of an "uncertainty awareness" mechanism. The core of this mechanism is the introduction of a weighted loss function for difficult regions. During training, the model not only learns segmentation accuracy but also learns to identify which regions are "difficult." The loss function is designed to assign higher weights to these difficult regions, causing the model to devote more attention and computational resources to correcting prediction errors in these regions during optimization. The underlying logic of this design is that by actively amplifying error signals, the model is forced to continuously iterate and optimize its understanding of complex boundaries during the learning process, thereby improving overall segmentation performance.

[0022] The basic cross-entropy loss function is: L_ce_{i,j} = - [G_{i,j} log(p_{i,j}) + (1-G_{i,j}) log(1-p_{i,j})] Where G is the ground truth and p is the probability that the model predicts whether the pixel is a gland.

[0023] The weighted loss function for the difficult region is: L_uce = (1 / (H W)) Σ_{i,j} [ Weight_{i,j} L_ce_{i,j} ] Weight_{i,j} = (1 - max(p, 1-p))^γ Here, γ (gamma) > 0 is used to control the intensity of weight amplification.

[0024] According to the formula, when p is closer to 0 or 1, it indicates that the model prediction has high certainty. At this time, 1 - max(p, 1-p) is close to 0 and Weight_{i,j} is small. When p is close to 0.5, it indicates that the model prediction has high uncertainty. At this time, Weight_{i,j} is large.

[0025] Two parallel feature extraction paths focus on spatial and temporal information respectively, and achieve deep collaboration and fusion between the two through a cross-attention mechanism.

[0026] Spatial Path: This path takes the segmented prostate gland region as input and employs a lightweight CNN-Transformer hybrid architecture. Its purpose is to finely extract static spatial structural features within the gland, such as tissue texture, boundary clarity, and morphological information of internal anomalous regions. This path ensures that the model does not lose crucial local details.

[0027] Temporal Path: This path takes the entire ultrasound video as input and employs a visual Transformer specifically designed for sequential data. Its core task is to model the dynamic evolution of the prostate gland over time in the video, such as subtle changes in gland morphology during scanning, dynamic patterns of blood flow, and the appearance and disappearance of abnormal echogenic areas.

[0028] The two paths do not operate in isolation, but are closely linked through a carefully designed dual-path attention fusion mechanism. This mechanism comprises two core components: 1. Intra-Attention: This operates within both the spatial and temporal paths to enhance the expressive power of features within each path. Intra-Attention in the spatial path strengthens the structural relationships between different regions of the gland, while Intra-Attention in the temporal path captures the dynamic relationships between different frames in the video sequence.

[0029] 2. Cross-Attention: This is the crucial bridge connecting two paths. It allows the static structural features of the spatial path to guide the temporal path to focus on specific keyframes in the video, and vice versa. For example, the spatial path might identify a suspicious structure in a specific region of a gland (such as the left peripheral zone), and then, through the cross-attention mechanism, guide the temporal path to focus on analyzing the dynamic changes of that region at different points in the video sequence. Similarly, the temporal path can also feed back any dynamic anomalies it discovers (such as a brief hyperecho in a frame) to the spatial path, guiding it to perform more refined spatial feature analysis.

[0030] Based on the fused comprehensive features, the system employs a lightweight classifier or regression model to make a final judgment on the salience of prostate cancer. The output can be a binary classification (e.g., "significant cancer" or "non-significant cancer") or a continuous risk score, i.e., a final diagnostic score (e.g., a probability value from 0 to 1). The specific steps are as follows: Similar to Grad-CAM, we first calculate the gradient of the model's final classification output relative to its input image. This gradient information reflects which regions in the image have the greatest impact on the final diagnosis. We extract the self-attention scores from the last Transformer layer of the dual-attention mechanism. These scores quantify the degree of mutual attention between different pixels (or regions) when the model processes the input, serving as the native "language" of the model's decision-making process. To eliminate scale differences between different attention heads, we perform L2 normalization on these scores. We then weight and fuse the normalized self-attention scores with the classification layer gradient to obtain a classification heatmap. The core logic of this step is that the self-attention scores provide the model's global attention structure in the input space, while the gradient provides the weights of how this attention structure influences the final classification decision. By combining the two, the generated heatmap not only highlights the regions the model focuses on but also accurately reflects the extent to which these regions contribute to the final diagnosis.

[0031] The prostate gland segmentation results, keyframe sequences, and final diagnostic scores are displayed on the interface in a graphic and textual format. Doctors can intuitively see the AI's analysis and "understanding" of the video. Through attention-guided interpretability, high-fidelity class activation heatmaps are generated, which intuitively show which specific areas and dynamic changes in the video the model focuses on when making a diagnosis. Doctors can zoom in and compare these classification heatmaps through the interactive interface to gain a deeper understanding of the AI's decision-making logic.

[0032] In summary, the steps of the embodiments of the present invention are as follows: The prostate gland segmentation model is used to segment the prostate region in each frame; simultaneously, keyframes in the video sequence are identified (and also segmented into glands); the prostate segmentation results of the keyframes are input into a dual-path feature extraction module to extract fused features. Based on the fused features, a classifier is used to classify salient / non-salient cancer to obtain the final diagnostic score. A classification heatmap is displayed, along with the previous gland segmentation results and intermediate results such as the keyframe sequence. The final result is the diagnostic score; other parameters are used for interpretability.

[0033] This invention also provides a prostate ultrasound video diagnostic system, which employs the methods described above. The system includes: Data is the cornerstone of AI models, and the quality and standardization of data directly determine the success or failure of subsequent analysis. The core task of this module is to perform automated and standardized preprocessing of raw ultrasound video data, providing high-quality input for the entire system.

[0034] To address the inherent complexity of ultrasound video data, an automated preprocessing pipeline was designed to integrate the data access and preprocessing modules.

[0035] The data access and preprocessing module uses the ECC image alignment algorithm to denoise each frame of the video to eliminate artifacts introduced by equipment, operation, or patient movement. It also uses techniques such as histogram equalization to standardize and enhance the images, ensuring that the images in the input model are consistent in terms of brightness and contrast. Based on preset algorithms and clinical expert experience, it performs quality assessment on the video sequence, automatically identifies and removes low-quality frames caused by equipment failure, image blurring, or improper operation, and retains only clear and complete video segments that can be analyzed.

[0036] Through this dual mechanism of "automation" and "standardization," this module not only provides stable and reliable input for subsequent AI analysis, but also solves the model generalization problem caused by data heterogeneity at its source, laying a solid foundation for the entire system to be used in multi-center, large-scale clinical applications.

[0037] If the data preprocessing module is the "entry point" of the system, then the intelligent processing module is the "technical heart" of the entire system. It carries the core algorithm of this invention and is responsible for transforming standardized raw ultrasound video data into deep medical features that can be used for decision-making through end-to-end AI reasoning.

[0038] The workflow for designing the intelligent processing module is a highly collaborative and automated analysis chain.

[0039] The intelligent processing module utilizes an advanced Transformer-based encoder-decoder architecture to accurately segment the prostate glands in each frame of the video, generating a high-quality segmentation mask. This enables prostate gland segmentation based on uncertainty perception, not only to locate the anatomical boundaries of the prostate but also to extract its spatial morphological features, providing precise anatomical context for subsequent analysis. After gland segmentation, the entire video is dynamically analyzed. By designing specialized attention mechanisms or feature clustering algorithms, keyframe sequences containing key anatomical structures, abnormal morphological changes, or important hemodynamic information of the prostate are intelligently identified. This step aims to extract the most diagnostically valuable information from the lengthy video, significantly improving the efficiency of subsequent analysis. Two parallel feature extraction paths are designed: one focuses on analyzing the static spatial structural features of the prostate glands, while the other aims to capture the dynamic temporal evolution features of the glands throughout the video. Through internal and cross-attention mechanisms, the two paths work in synergy, ultimately fusing to generate a comprehensive deep feature vector that fully characterizes the prostate state and contains both spatial and temporal information.

[0040] Therefore, the intelligent processing module is the bridge connecting the raw data and the final diagnostic decision. Its high-quality, multi-dimensional output is the fundamental basis for the judgment decision module to make accurate judgments.

[0041] After the intelligent processing module extracts deep medical features, the discrimination and decision-making module becomes the "intelligent hub" of the system. Its core task is to comprehensively analyze and integrate these complex features, and finally generate diagnostic results that can be directly used by clinicians.

[0042] The decision-making module is used to make a final judgment on the significance of prostate cancer based on the fused comprehensive features. The system employs a lightweight classifier or regression model. The output can be a binary classification (such as "significant cancer" or "non-significant cancer") or a continuous risk score (such as a probability value from 0 to 1).

[0043] The ultimate value of the technology lies in the convenience and effectiveness of its clinical application. The clinical interaction module is the "user interface" connecting the AI ​​system of this invention with clinicians. Its core design goal is to achieve human-machine collaboration, presenting complex AI analysis results in an intuitive and user-friendly way, thereby seamlessly integrating into the doctor's daily workflow. This module provides doctors with an integrated and interactive work platform.

[0044] The clinical interaction module displays the prostate gland segmentation results generated by the intelligent processing module, the identified keyframe sequences, and the final diagnostic score given by the decision-making module in a graphic and textual format on the interface. Doctors can intuitively see the AI's analysis and "understanding" of the video. Through attention-guided interpretability, high-fidelity class activation heatmaps are generated, which intuitively show which specific areas and dynamic changes in the video the model focuses on when making a diagnosis. Doctors can zoom in and compare these heatmaps through the interactive interface to gain a deeper understanding of the AI's decision-making logic.

[0045] The four modules of this invention are interconnected and together constitute an end-to-end intelligent diagnostic pipeline, ensuring the integrity, efficiency and clinical applicability of the technical solution.

[0046] This invention provides a method and system for prostate ultrasound video diagnosis. It features a fully automated information extraction module capable of accurately segmenting the prostate gland region from raw, non-standardized ultrasound video streams and intelligently identifying dynamic frame sequences containing key diagnostic information. This eliminates over-reliance on expert experience and enables efficient and accurate extraction of key information from ultrasound videos. The embodiment also employs an advanced feature fusion strategy that organically integrates fine spatial features from the prostate region, global temporal dynamic features of the entire video stream, and key diagnostic features from key frames. This constructs a comprehensive and three-dimensional patient image representation, effectively fusing multi-dimensional features from the prostate region, the entire video stream, and key frames. Furthermore, this embodiment develops a high-fidelity, high-precision interpretability module that intuitively demonstrates to doctors which specific regions and dynamic changes in the video the model focused on when making a diagnosis. This enhances the model's transparency and clinical reliability, improving interpretability and gaining the trust of clinicians.

[0047] This invention encapsulates three interrelated and indispensable core innovations. Together, they form a complete technological closed loop from data input to diagnostic output and result interpretation, aiming to achieve a disruptive breakthrough in AI-assisted diagnosis of prostate cancer, as detailed below: 1. End-to-End Intelligent Diagnostic Pipeline: This invention pioneers a fully automated, end-to-end intelligent diagnostic pipeline, from raw ultrasound video data input to final cancer saliency prediction output. This pipeline integrates multiple modules, including data access, automatic preprocessing, intelligent analysis, and decision output, achieving a high degree of automation in the diagnostic process. This innovation not only significantly improves diagnostic efficiency, but more importantly, it eliminates biases introduced by manual operation through standardized processes, laying a solid foundation for subsequent accurate feature analysis and model training.

[0048] 2. Dual-Path Attention Fusion Model Architecture: Addressing the challenges of multi-dimensional feature fusion, this invention innovatively designs a dual-path attention fusion model architecture. This architecture comprises two parallel feature extraction paths: one focuses on capturing the fine spatial structural features of the prostate region, while the other dedicates itself to modeling the global temporal dynamic evolution of the entire video. By introducing carefully designed internal attention and cross-attention mechanisms, the model can achieve deep collaborative learning and fusion of spatial and temporal information. This architecture systematically solves for the first time the problem of how to simultaneously extract and effectively combine static spatial and dynamic temporal information from ultrasound videos, providing strong technical support for accurate diagnosis.

[0049] 3. Attention-Guided Interpretability Module: To overcome the "black box" problem of AI models, this invention develops a high-fidelity interpretability module based on attention guidance. This module directly utilizes the self-attention scores generated during the inference process of the dual-path attention fusion model, combined with classification layer gradient information, to generate a class activation heatmap that accurately reflects the model's decision-making basis. This method, for the first time, transforms the complex attention mechanism within the model into a visually understandable explanation for clinicians, clearly demonstrating the abnormal prostate region that the model focuses on, thereby significantly improving the model's interpretability and paving the way for its widespread adoption in clinical practice.

[0050] In summary, the three major innovations of this invention are not isolated technical improvements, but rather a tightly coupled organic whole. The end-to-end pipeline ensures the standardization and efficiency of data processing, the dual-path attention fusion model is the core algorithm engine for achieving accurate diagnosis, and the attention-guided interpretability module is a key bridge connecting technological innovation and clinical trust. These three elements work synergistically to construct an efficient, accurate, and reliable AI-assisted diagnostic system for prostate cancer.

[0051] This invention systematically proposes an end-to-end intelligent diagnostic pipeline from raw ultrasound video data to final cancer significance prediction, aiming to address core clinical challenges in current prostate cancer diagnosis processes, such as invasiveness, subjectivity, and high false negative rates. This pipeline achieves end-to-end automation of data processing, intelligent analysis, decision generation, and clinical interaction through the collaborative work of four modules.

[0052] Its core technology lies in three interrelated innovations. First, the prostate gland segmentation module based on uncertainty perception, through an innovative loss function design, significantly improves the segmentation accuracy and robustness of gland contours in complex clinical images, laying a solid foundation for subsequent analysis. Second, the spatiotemporal video understanding module based on a dual-path attention mechanism innovatively constructs two parallel feature extraction paths—space and time—and through a carefully designed attention fusion mechanism, enables the model to more comprehensively and deeply understand the static structure and dynamic evolution information of the prostate region, thereby achieving efficient utilization of video information. Finally, the final classification module based on multi-instance learning simulates the logic of clinicians making decisions by integrating evidence from multiple frames, effectively improving the model's generalization ability and diagnostic accuracy on small-sample medical data.

[0053] Crucially, the embodiments of this invention do not stop at performance improvement. Through an attention-guided interpretability module, we have successfully transformed complex deep learning models from "black boxes" to "white boxes." The high-fidelity heatmaps generated by this module intuitively reveal the key areas that the model focuses on when making diagnoses, providing clinicians with powerful decision-making evidence and verification tools, and greatly enhancing the clinical reliability of the AI ​​system.

Claims

1. A method for prostate ultrasound video diagnosis, characterized in that, Includes the following steps: Image compensation was performed on each frame of the prostate ultrasound video for inter-frame motion caused by heartbeat, respiration and probe movement, and temporal smoothing filtering was performed. The quality of the processed prostate ultrasound video was then evaluated to obtain the preprocessed prostate ultrasound video. By utilizing a Transformer-based encoder-decoder structure and introducing a weighted loss function for difficult regions, multi-scale spatial morphological features in the preprocessed prostate ultrasound video are captured and given higher weights to enhance uncertainty perception. The multi-scale spatial features are mapped back to the original image space to generate a pixel-level segmentation mask for the prostate gland. The pixel-level segmentation mask for the prostate gland and the preprocessed prostate ultrasound video are input into a dual-path attention fusion model to extract information in both spatial and temporal dimensions in parallel and perform deep collaboration and fusion. Finally, keyframe sequences are identified and fused to generate a comprehensive deep feature vector that represents the prostate state and contains both spatial and temporal information. The significance of prostate cancer is finally determined based on comprehensive deep feature vectors, and a final diagnostic score is output.

2. The method for prostate ultrasound video diagnosis as described in claim 1, characterized in that, The process of inputting pixel-level segmentation masks of the prostate gland and preprocessed prostate ultrasound video into a dual-path attention fusion model to extract spatial and temporal information in parallel and perform deep collaboration and fusion includes the following steps: The pixel-level segmentation mask of the prostate gland is input into the spatial path. With internal attention to enhance the structural relationship between different regions of the gland, the static spatial structure features inside the gland are extracted in detail. The preprocessed prostate ultrasound video is input into the time path, and the dynamic temporal evolution characteristics of the prostate gland in the video are modeled by capturing the dynamic correlation between different frames in the video sequence through internal attention. If the spatial path identifies a suspicious structure in a specific region, then cross-attention guides the temporal path to focus on analyzing the dynamic changes of that region at different time points in the video sequence. If the time path identifies a dynamic anomaly at a specific time, then cross-attention guides the spatial path to focus on identifying and extracting static spatial structure features. Finally, keyframe sequences were identified and fused to generate a comprehensive deep feature vector that represents the prostate state and contains both spatial and temporal information.

3. A prostate ultrasound video diagnostic system, characterized in that, The system employing the prostate ultrasound video diagnostic method as described in claim 1 comprises: The data access and preprocessing module is used to perform artifact denoising, brightness and contrast improvement on each frame of the prostate ultrasound video, perform image compensation for inter-frame motion caused by heartbeat, breathing and probe movement, and perform temporal smoothing filtering. It also performs quality assessment on the processed prostate ultrasound video, identifies and removes low-quality frames caused by equipment failure, image blur or improper operation, and obtains the preprocessed prostate ultrasound video. The intelligent processing module utilizes a Transformer-based encoder-decoder structure, introduces a weighted loss function for difficult regions, captures multi-scale spatial morphological features in the preprocessed prostate ultrasound video, assigns them higher weights to enhance uncertainty perception, maps the multi-scale spatial features back to the original image space, generates a pixel-level segmentation mask for the prostate gland, and inputs the pixel-level segmentation mask and the preprocessed prostate ultrasound video into a dual-path attention fusion model. It extracts information in the spatial and temporal dimensions in parallel and performs deep collaboration and fusion, finally identifying key frame sequences and fusing them to generate a comprehensive deep feature vector that represents the prostate state and contains both spatial and temporal information. The discrimination decision module is used to make a final judgment on the significance of prostate cancer based on a comprehensive deep feature vector and output the final diagnostic score.