A medical imaging lesion detection system

Abstract

The present invention relates to a technical field, specifically a medical imaging lesion detection system, the system includes a data processing module, a quality control module, a resolution reconstruction module, a data enhancement module, a dynamic learning module, a real-time feedback module, a model interpretability enhancement module, and a comprehensive diagnosis module. The present invention achieves the improvement of image resolution by optimizing the quality of the image, and combines the conditional generative adversarial network to simulate the lesion characteristics in the medical image, which significantly improves the efficiency and accuracy of the medical imaging lesion detection system. The improvement of image resolution ensures the high quality and authenticity of the image, and provides accurate basic data for lesion detection. The application of the conditional generative adversarial network expands the data set and enhances the model's ability to identify rare lesions. The dynamic learning module allows the model to self-optimize according to the doctor's feedback, which improves the adaptability and diagnostic accuracy. The intervention of the real-time feedback module enables the operator to obtain quality feedback in the diagnosis process in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of image analysis, and in particular to a medical image lesion detection system. Background Art

[0002] The field of image analysis technology focuses on processing, analyzing and interpreting image content through computer vision and machine learning algorithms, aiming to extract useful information from images to support decision making. Image analysis technology is widely used in multiple industries, among which the application in the medical field is particularly important, involving key aspects such as early detection of diseases, assistance in the diagnosis process and evaluation of treatment effects. Through in-depth analysis of medical imaging data, this technology can identify the differences between normal and abnormal structures, thereby providing doctors with accurate diagnostic information.

[0003] Among them, the medical image lesion detection system is a system that applies image analysis technology to the medical field. Its main purpose is to identify and classify lesion areas in medical images, including tumors, inflammation or other abnormal tissues, through automated detection methods. The system aims to achieve rapid and accurate diagnosis, assist doctors in disease assessment, and thus significantly improve the efficiency and accuracy of medical diagnosis. By detecting diseases at an early stage, the system helps to achieve early treatment, improve efficacy and patient survival rate.

[0004] Traditional systems face limitations in image quality and resolution when processing medical images, which directly affects the accuracy of lesion detection and the potential for early diagnosis. The lack of sufficient data enhancement means and dynamic learning capabilities limits the effectiveness in dealing with complex or rare lesions. The lack of effective feedback mechanisms and explanatory support makes it difficult for doctors to evaluate diagnostic results, affecting the acceptance of diagnoses and clinical decision-making. The lack of real-time feedback and direct involvement of operators results in the inability to promptly discover and correct problems that arise during the diagnostic process, affecting diagnostic efficiency and accuracy. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a medical imaging lesion detection system.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a medical imaging lesion detection system, the system comprising:

[0007] The data processing module adjusts the brightness of the image based on the original medical image, optimizes the random noise and image edges in the image, and generates a processed image;

[0008] The quality control module identifies blurry areas and artifacts in the image based on the processed image, and uses image enhancement technology to refine the image quality and generate a quality-optimized image;

[0009] The resolution reconstruction module optimizes the image resolution based on the quality optimized image, determines the image authenticity, and iteratively optimizes the network parameters to generate a resolution adjusted image;

[0010] The data enhancement module adjusts the image based on the resolution, uses a conditional generative adversarial network to simulate the lesion characteristics in the medical image, constructs a synthetic image of the target feature, and expands the original data set in combination with the synthetic image to generate an enhanced data set;

[0011] The dynamic learning module adjusts the model parameters based on the enhanced data set and according to the doctor's feedback as a reward signal, optimizes the model error, and generates an optimized recognition model;

[0012] The real-time feedback module analyzes the collected images in real time based on the optimized recognition model, identifies image quality problems, and feeds back to the operator through the user interface to generate real-time optimization feedback information;

[0013] The model explanation enhancement module applies a multi-head attention mechanism based on the optimized recognition model, analyzes the image area of ​​the model target, visualizes the area, highlights the model focus, and generates an explanation enhancement model;

[0014] The comprehensive diagnosis module analyzes the lesion area based on the explanatory enhancement model and the real-time optimization feedback information, and organizes and formats the diagnosis data to generate lesion diagnosis information.

[0015] The improvements of the present invention are that the processed image includes an adjusted brightness level, an adjusted noise level and an optimized image edge, the quality optimized image includes a corrected blurred area, an eliminated artifact and refined image details, the enhanced data set includes a synthesized lesion feature image, an expanded number of samples, the optimized recognition model includes an adjusted learning rate, an optimized loss function and a solidified recognition strategy, the explanatory enhancement model includes a visualized area of ​​interest, annotated key feature points and an optimized model resolution, and the lesion diagnosis information includes an analyzed lesion area, a summarized diagnosis logic and formatted diagnosis information content.

[0016] The present invention is improved in that the data enhancement module comprises:

[0017] The conditional GAN ​​development submodule adjusts the image based on the resolution, adopts a conditional generative adversarial network, sets the generator to receive the lesion label as a conditional input, combines the image and lesion information, evaluates the authenticity of the synthesized image and the accuracy of the lesion feature, and generates a lesion feature synthesized image;

[0018] The synthetic image quality assessment submodule calculates the Laplace variance of the image based on the lesion feature synthetic image to perform accuracy assessment, verifies the matching of the quality of the synthetic image with the analysis requirements, and generates a synthetic image assessment record;

[0019] The data enhancement strategy optimization submodule rotates, scales and crops the image based on the synthetic image evaluation record to optimize the diversity of the data set, and expands the original data set in combination with the obtained synthetic image to generate an enhanced data set.

[0020] The present invention is improved in that the conditional generative adversarial network is based on the formula:

[0021] ;

[0022] Calculate the loss of the discriminator;

[0023] in, and To adjust the weights, and is the regularization coefficient, For real images, is the lesion label, is a random noise input, and is the probability of the discriminator evaluating real and synthetic images, and is the output of the discriminator on low-level features for real and synthetic images.

[0024] The present invention is improved in that the dynamic learning module comprises:

[0025] The deep Q network training submodule performs deep Q network training based on the enhanced data set, including matching the input layer with the image data size, extracting key features in the medical image, and optimizing the training process to generate a preliminary recognition model;

[0026] The doctor feedback integration submodule integrates the doctor's diagnosis feedback as a reward signal based on the preliminary recognition model, adjusts the model learning target through positive and negative feedback, and matches the model to the medical diagnosis standard to generate a feedback adjustment model;

[0027] The recognition strategy adjustment submodule adjusts the model based on the feedback, analyzes the doctor's feedback and model performance, identifies the model's weaknesses in identifying the target lesions, optimizes the accuracy of identifying the target lesions, matches it to the doctor's diagnostic ideas, and generates an optimized recognition model.

[0028] The present invention is improved in that the real-time feedback module comprises:

[0029] The real-time image analysis submodule performs image processing based on the optimized recognition model, including adjusting the image size and applying filters to optimize image features, and uses the processed image to perform lesion recognition tasks to generate real-time analysis data;

[0030] The quality assessment model submodule assesses the quality parameters of the image based on the real-time analysis data, scores the image quality according to the model output, identifies image quality problems according to the score, and generates a quality assessment result;

[0031] The optimization strategy proposal submodule formulates an image optimization strategy based on the quality assessment results and the identified image quality issues, optimizes the image quality, and feeds back to the operator through the user interface to generate real-time optimization feedback information.

[0032] The present invention is improved in that the model explanation enhancement module comprises:

[0033] The attention mechanism integration submodule is based on the optimized recognition model and adopts a multi-head attention mechanism to optimize the model's attention to key image areas by calculating the weight distribution between input features, and adjusts the model's sensitivity to key parts of the image to generate a focus image;

[0034] The feature visualization submodule uses gradient weighted class activation mapping based on the focus image to highlight the key areas of model decision, generates a heat map through a weighted combination of gradient and feature map, and processes the superposition of the heat map and the original image to generate a visual explanation map;

[0035] The explanatory feedback loop submodule adjusts the attention mechanism based on the visual explanation graph and the doctor's feedback, analyzes the doctor's feedback on the visual output, identifies the deficiencies in the model analysis, and iterates repeatedly to generate an explanatory enhancement model.

[0036] The present invention improves the multi-head attention mechanism according to the formula:

[0037] ;

[0038] Calculate weighted output values ​​to highlight key image areas;

[0039] in, is the query matrix, is the key matrix, is the value matrix, is the context relevance matrix, is the dimension of the key vector, is a small amount, used to avoid division by zero errors, Represents the query matrix With key matrix The transpose of is used for dot product operation, Represents the multi-head self-attention function.

[0040] Compared with the prior art, the advantages and positive effects of the present invention are:

[0041] In the present invention, the image resolution is improved by optimizing the quality of the image, and the conditional generative adversarial network is combined to simulate the lesion characteristics in the medical image, which significantly improves the efficiency and accuracy of the medical image lesion detection system. The improvement of image resolution ensures the high quality and authenticity of the image, and provides accurate basic data for lesion detection. The application of the conditional generative adversarial network expands the data set and enhances the model's ability to recognize rare lesions. The dynamic learning module allows the model to self-optimize according to the doctor's feedback, which improves the adaptability and diagnostic accuracy. The intervention of the real-time feedback module enables the operator to obtain quality feedback in the diagnosis process in real time, which enhances the interactivity and real-time nature of the diagnosis process. The interpretability enhancement module improves the interpretability of the model through visualization processing, so that the doctor can intuitively understand the basis for the model's judgment and increase the confidence in the diagnosis. The comprehensive diagnosis module makes the diagnostic information of the lesion area more accurate, and promotes the clarity of the diagnostic logic and the standardization of operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A module diagram of a medical imaging lesion detection system is proposed for the present invention;

[0043] Figure 2 A system framework diagram of a medical imaging lesion detection system is proposed for the present invention;

[0044] Figure 3 A flow chart of a data processing module in a medical imaging lesion detection system is proposed for the present invention;

[0045] Figure 4 A flow chart of a quality control module in a medical imaging lesion detection system is proposed for the present invention;

[0046] Figure 5 The present invention proposes a flow chart of a resolution reconstruction module in a medical imaging lesion detection system;

[0047] Figure 6 The present invention proposes a flow chart of a data enhancement module in a medical imaging lesion detection system;

[0048] Figure 7 A flow chart of a dynamic learning module in a medical imaging lesion detection system is proposed for the present invention;

[0049] Figure 8 The present invention proposes a flow chart of a real-time feedback module in a medical imaging lesion detection system;

[0050] Fig. 9A flow chart of a model interpretability enhancement module in a medical imaging lesion detection system is proposed for the present invention;

[0051] Fig.10 The present invention proposes a flow chart of a comprehensive diagnosis module in a medical imaging lesion detection system. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0053] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating positions or positional relationships, are based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0054] Example: See Figure 1 The present invention provides a technical solution: a medical imaging lesion detection system comprising:

[0055] The data processing module adjusts the brightness of the image based on the original medical image, optimizes the random noise and image edges in the image, and generates a processed image;

[0056] The quality control module analyzes image features based on processed images, identifies blurred areas and artifacts in the images, and uses image enhancement technology to perform regional trimming, refine image quality, and generate quality-optimized images.

[0057] The resolution reconstruction module optimizes the image resolution based on the quality optimized image, determines the image authenticity, and iteratively optimizes the network parameters through repeated training processes to generate resolution adjusted images;

[0058] The data enhancement module adjusts images based on resolution, uses conditional generative adversarial networks to simulate lesion features in medical images, adjusts the model through the training process, constructs synthetic images of target features, and expands the original data set with the synthetic images to generate an enhanced data set.

[0059] The dynamic learning module adjusts the model parameters based on the enhanced data set and the doctor's feedback as a reward signal, performs iterative model training, optimizes the model's error, and adjusts and solidifies the model's recognition strategy to generate an optimized recognition model.

[0060] The real-time feedback module analyzes the collected images in real time based on the optimized recognition model, identifies image quality issues, and feeds back to the operator through the user interface to generate real-time optimization feedback information;

[0061] The model interpretation enhancement module is based on the optimized recognition model and applies the multi-head attention mechanism to analyze the image area of ​​the model target. It combines the gradient weighted class activation mapping to visualize the area, highlight the model focus, and generate an interpretation enhancement model.

[0062] The comprehensive diagnosis module analyzes the lesion area based on the explanatory enhancement model and real-time optimization feedback information, organizes and formats the diagnostic data in combination with the model output and diagnostic logic, and generates lesion diagnosis information.

[0063] The processed images include adjusted brightness levels, adjusted noise levels and optimized image edges. The quality-optimized images include corrected blurred areas, eliminated artifacts and refined image details. The enhanced datasets include synthesized lesion feature images and expanded sample numbers. The optimized recognition model includes adjusted learning rates, optimized loss functions and solidified recognition strategies. The explanatory enhancement model includes visualized areas of interest, annotated key feature points and optimized model resolution. The lesion diagnosis information includes analyzed lesion areas, summarized diagnostic logic and formatted diagnostic information content.

[0064] See also Figure 2 and Figure 3 ,The data processing module includes a histogram equalization submodule, a Gaussian filtering denoising submodule, and an edge enhancement submodule;

[0065] The histogram equalization submodule performs image processing based on the original medical image. It calculates the histogram of the image, redistributes pixels, and equalizes the brightness distribution of the image to generate a brightness-optimized image.

[0066] The histogram equalization submodule collects the original medical image data, calculates the frequency of occurrence of each pixel gray value, and forms a histogram. Then, the cumulative distribution function (CDF) is used to adjust the pixel value according to the histogram to make the brightness evenly distributed, thereby improving the image contrast. To ensure the processing accuracy, the linear interpolation method is used to calculate the new gray value of each pixel. Through the implementation steps, the brightness of medical images can be effectively optimized, the visual effect can be improved, and the efficiency of subsequent processing can be enhanced.

[0067] The Gaussian filter denoising submodule optimizes the image based on brightness, smoothes the image, and automatically calculates the image to optimize the randomly distributed noise points in the image to generate a denoised image;

[0068] The Gaussian filter denoising submodule first defines a Gaussian filter kernel. The kernel size and standard deviation are determined through experimental data to meet the needs of different image details. The filter kernel is applied to the brightness optimized image, and each pixel is processed by convolution operation. The core operation is to weighted average the grayscale value of each pixel with its surrounding pixels. The weight is determined by the Gaussian function. This can effectively reduce the impact of noise while maintaining image details. After calculation and comparison, it is ensured that the denoising effect meets the expected standard.

[0069] The edge enhancement submodule performs edge recognition on the image based on the denoised image, identifies the edges and noise in the image, and highlights the key edges of the image to generate a processed image;

[0070] The edge enhancement submodule first applies the Sobel operator to calculate the gradient strength and direction of the image. Areas with higher gradient strength are marked as edges. After that, the contrast of the edge area is enhanced and the interference of the non-edge area is reduced. Each step of the calculation is accurate to the pixel level to ensure that the edges are clearly visible. After completing the steps, the output image will have more obvious edges, providing doctors with clearer diagnostic images.

[0071] See also Figure 2 and Figure 4 ,The quality control module includes an image quality assessment submodule, an instant feedback system submodule, and an automatic adjustment mechanism submodule;

[0072] The image quality assessment submodule builds a network structure including residual blocks and attention based on the processed images, performs feature distribution analysis on the key features of the image, identifies blurred areas and artifacts in the image, and generates quality assessment records;

[0073] The image quality assessment submodule is based on processed images. It first collects the images to be analyzed and extracts the original image data, including pixel distribution, color depth and brightness level. Using these data, it builds a deep network model with residual blocks and attention mechanism to analyze the key features of the image. By calculating the residuals of each area in the image and combining the attention mechanism to evaluate the importance of each area, it can identify the blurred areas and artifacts in the image and generate a comprehensive quality assessment record with these data.

[0074] The instant feedback system submodule plans image correction strategies and generates data correction plans based on quality assessment records by analyzing key indicators and problem areas in image quality data;

[0075] The instant feedback system submodule is based on quality assessment records. It first compares the data in the records with the image quality standards to identify key indicators and specific problem areas that do not meet the standards. By analyzing the characteristic distribution and deviation from the standards of these problem areas, it plans specific image correction strategies, including correction of color distortion, contrast adjustment and sharpening, thereby generating a detailed data correction plan.

[0076] The automatic adjustment mechanism submodule performs local quality adjustment of the image based on the data correction plan. By setting the filtering parameters, the image is thresholded, the quality problem area of ​​the image is refined, and the quality optimized image is generated;

[0077] The automatic adjustment mechanism submodule is based on the data correction plan. It first analyzes the correction measures in the plan and sets the corresponding image processing parameters according to the specific content of each measure. The parameters include filter intensity, threshold setting and area selection criteria. By precisely controlling the parameters, the image is locally adjusted in terms of quality. The image quality problem areas are refined mainly through dynamic threshold processing and regional key optimization. During the adjustment process, the adjustment effect is monitored in real time, and the processing parameters are adjusted based on the effect feedback to generate an image with optimized quality.

[0078] See also Figure 2 and Figure 5 ,The resolution reconstruction module includes the SRGAN training submodule, the loss function customization submodule, and the multi-scale feature fusion submodule;

[0079] The SRGAN training submodule trains a super-resolution generative adversarial network model based on quality-optimized images, builds a generator and a discriminator, optimizes the resolution and veracity of images, and adjusts image details to generate resolution-enhanced images.

[0080] The SRGAN training submodule is based on quality-optimized images. First, the optimized images are collected as training data and the data is preprocessed, including image normalization and size unification. After that, a super-resolution generative adversarial network (SRGAN) consisting of a generator and a discriminator is constructed. The generator is responsible for enhancing the resolution of the image through deep learning technology, while the discriminator evaluates the authenticity of the generated image and ensures the naturalness of the image details. During the training process, the parameters of the generator and the discriminator are continuously optimized through repeated iterations to improve their ability to process image details. Finally, through this training process, images with higher resolution and improved details are generated. The parameter configuration and results of each step of training are recorded to provide data support for subsequent model verification and application.

[0081] The loss function customization submodule adjusts the loss function of the super-resolution generative adversarial network model based on the resolution-enhanced image, calculates the image content relevance, evaluates the difference between the generated image and the real image through adversarial loss, and adjusts the weight parameters in the loss function to generate a loss function optimized image;

[0082] The loss function customization submodule is based on the resolution-enhanced image. It first analyzes the content correlation between the enhanced image and the original image, and uses content loss to quantify the similarity between the two. At the same time, adversarial loss is added to evaluate the difference in visual authenticity between the generated image and the real image. On this basis, the weight parameters in the loss function are adjusted, such as increasing the weight of content loss to strengthen the model's emphasis on image content fidelity, and adjusting the weight of adversarial loss to improve the overall visual effect of the image. In this way, the loss function is optimized to make it more in line with the image quality requirements in practical applications. Finally, the adjusted loss function is generated and applied to the SRGAN model to generate images with optimized loss evaluation.

[0083] The multi-scale feature fusion submodule optimizes the image based on the loss function and uses the feature pyramid network to optimize the super-resolution generative adversarial network model to capture image features at different scales, optimize the expression of image details and global information, and generate resolution-adjusted images.

[0084] The multi-scale feature fusion submodule optimizes the image based on the loss function and introduces a feature pyramid network to enhance the model's feature extraction and fusion capabilities. By designing a multi-layer network structure, each layer is responsible for capturing image features at different scales, from subtle textures to macro structures. Then, these features are fused layer by layer to optimize the image's detail performance and the consistency of global information. During the feature fusion process, the weight distribution of features at each layer is adjusted to ensure that important features can be highlighted and secondary features will not interfere with the main visual effects. Through this hierarchical and weighted feature processing, the effect of the super-resolution generative adversarial network is optimized, and ultimately a resolution image with fine-tuned adjustments is generated. The configuration parameters and effect evaluation of feature fusion are recorded to provide a technical reference for subsequent image processing.

[0085] See also Figure 2 and Figure 6 ,The data enhancement module includes the conditional GAN ​​development submodule, the ,synthetic image quality assessment submodule, and the data enhancement ,strategy optimization submodule;

[0086] The conditional GAN ​​development submodule adjusts the image based on the resolution, adopts the conditional generative adversarial network, sets the generator to receive the lesion label as the conditional input, combines the image and lesion information, evaluates the authenticity of the synthetic image and the accuracy of the lesion features, and generates a synthetic image with lesion features;

[0087] The conditional GAN ​​development submodule is designed to use the lesion label as a condition to control the generated image content so that the lesion features in the generated image meet the predetermined conditions. The lesion information is embedded into the image generation process through the encoder. The generator outputs a synthetic image with target lesion features in combination with the resolution-adjusted image. The discriminator adjusts the generation model based on the comparison between the synthetic image and the real image to improve the authenticity of the synthetic image and the accuracy of the lesion features, so that the synthetic lesion image gradually approaches the texture and details of the real image during the training process.

[0088] Conditional generation of adversarial networks, according to the formula:

[0089] ;

[0090] Calculate the loss of the discriminator;

[0091] in, and To adjust the weights, and is the regularization coefficient, For real images, is the lesion label, is a random noise input, and is the probability of the discriminator evaluating real and synthetic images, and is the output of the discriminator on low-level features for real and synthetic images.

[0092] Execution process:

[0093] First, set a reasonable weight coefficient and , determined by optimizing the adjustment period, and then applying and Lesion labeling and the discriminator on low-level features The error between the outputs is regularized with the squared norm to enhance the model's ability to learn and identify lesion features. Furthermore, the parameters of the generator and discriminator are adjusted through the gradient descent method to minimize the entire loss function, thereby improving the authenticity of the synthetic images and the accuracy of the lesion features.

[0094] The synthetic image quality assessment submodule synthesizes images based on lesion features, calculates the Laplace variance of the image for accuracy assessment, verifies the matching of the quality of the synthetic image with the analysis requirements, and generates synthetic image assessment records;

[0095] Synthetic image quality assessment submodule, Laplace variance is used as a tool to measure image clarity and detail richness to evaluate the quality of synthetic images. Image clarity is determined by calculating the edge strength in the image. A high Laplace variance value indicates that the image has high contrast and clear edges, which is especially important for medical images. The evaluation results will be recorded to provide a reference for subsequent data enhancement strategies.

[0096] The data enhancement strategy optimization submodule rotates, scales, and crops images based on synthetic image evaluation records to optimize the diversity of the data set, and combines the obtained synthetic images to expand the original data set to generate an enhanced data set;

[0097] The data enhancement strategy optimization submodule uses data enhancement technology to display lesion characteristics from multiple angles and scales, enhancing the model's generalization ability and its ability to recognize unseen lesion types. Rotation and scaling operations enable the model to recognize lesions with different directions and sizes, while cropping helps the model focus on learning local features of the image. The resulting dataset not only increases in quantity, but also has significantly improved quality and diversity, providing more accurate and comprehensive data support for medical diagnosis.

[0098] See also Figure 2 and Figure 7 ,The dynamic learning module includes the deep Q network training submodule, the ,doctor feedback integration submodule, and the recognition strategy adjustment submodule;

[0099] The deep Q network training submodule performs deep Q network training based on the enhanced data set, including matching the input layer to the image data size, extracting key features in medical images, and optimizing the training process to generate a preliminary recognition model;

[0100] The deep Q network training submodule automatically learns and extracts key features in medical images through deep learning methods. By utilizing the diverse images in the enhanced data set, the network is repeatedly trained to adjust internal parameters to minimize prediction errors. The input layer is designed to adapt to medical images of different sizes to ensure that information is not lost. By combining structures such as convolutional layers, pooling layers, and fully connected layers, the deep Q network can effectively identify and enhance key features in the image, such as the shape, size, and texture of the lesion area, thereby generating a preliminary model that can accurately identify features.

[0101] The doctor feedback integration submodule is based on the preliminary recognition model, integrates the doctor's diagnosis feedback as a reward signal, adjusts the model learning target through positive and negative feedback, and matches the model to the medical diagnosis standard to generate a feedback adjustment model;

[0102] In the physician feedback integration submodule, the physician's diagnostic feedback is used as a key signal to adjust and optimize the model. The physician evaluates the diagnostic results of the preliminary model and provides specific positive and negative feedback. Positive feedback strengthens the model's learning path when making a correct diagnosis, while negative feedback guides the model to adjust cases that were not correctly identified. In this process, the model's learning objectives are continuously updated to more accurately meet the actual standards of medical diagnosis. Through this dynamic adjustment, the accuracy and applicability of the model are gradually improved, and ultimately a model that can more accurately reflect the physician's diagnostic thinking is generated.

[0103] The recognition strategy adjustment submodule adjusts the model based on feedback, analyzes doctor feedback and model performance, identifies the model's weaknesses in identifying target lesions, optimizes the accuracy of identifying target lesions, matches it to the doctor's diagnostic ideas, and generates an optimized recognition model;

[0104] The recognition strategy adjustment submodule determines the strengths and weaknesses of the model by deeply analyzing the model's performance in identifying various types of lesions and the specific feedback from doctors. It makes targeted adjustments and optimizations to the weaknesses in the recognition process, such as inaccurate or confusing recognition of certain types of lesions. By improving the training algorithm and adjusting the network parameters, it improves the model's recognition ability and accuracy in complex or rare lesions, ensuring that the final generated model can fully meet the doctors' diagnostic needs, thereby improving the overall efficiency and accuracy of medical imaging diagnosis.

[0105] See also Figure 2 and Figure 8 ,The real-time feedback module includes a real-time image analysis submodule, a quality assessment model submodule, and an optimization strategy proposal submodule;

[0106] The real-time image analysis submodule performs image processing based on the optimized recognition model, including adjusting the image size and applying filters to optimize image features, and uses the processed images to perform lesion recognition tasks to generate real-time analysis data;

[0107] The real-time image analysis submodule first receives imaging data from various medical devices and automatically adjusts the image size to adapt to the input requirements of the recognition model. Subsequently, a series of filters, such as sharpening filters and edge detection filters, are applied to enhance the key features of the image and make the lesion area more obvious. The preprocessing step improves the diagnosability of the image. The optimized recognition model is used to quickly detect and classify lesions and generate real-time diagnostic data. In this process, processing speed and accuracy are key indicators to meet clinical needs.

[0108] The quality assessment model submodule evaluates the quality parameters of the image based on real-time analysis data, scores the image quality according to the model output, identifies image quality issues based on the scores, and generates quality assessment results;

[0109] The quality assessment model submodule uses image quality assessment algorithms, such as assessment techniques based on image texture and clarity, to give each image a comprehensive score. This score reflects the overall quality of the image, including key parameters such as contrast and sharpness. It automatically identifies poor quality images based on set thresholds and marks quality issues that may lead to diagnostic errors. The assessment results will be recorded in detail and used for subsequent optimization strategy formulation.

[0110] The optimization strategy proposal submodule formulates an image optimization strategy based on the quality assessment results and the identified image quality issues, optimizes the image quality, and feeds back to the operator through the user interface to generate real-time optimization feedback information;

[0111] The optimization strategy proposal submodule uses the problem areas pointed out in the quality assessment results to design specific image optimization measures, including contrast enhancement, noise suppression, and dynamic range adjustment. Each strategy aims to solve specific image quality problems, thereby improving the overall diagnostic effect of the image. The optimized results will be displayed to the operator in real time through the user interface, providing adjustment suggestions and operational guidance. This feedback mechanism ensures that the medical team can make diagnoses based on the highest quality images and improve the efficiency and safety of medical services.

[0112] See also Figure 2 and Fig. 9 ,The model interpretability enhancement module includes the attention mechanism integration submodule, the feature visualization submodule, and the interpretable feedback loop submodule;

[0113] The attention mechanism integration submodule is based on the optimized recognition model and adopts a multi-head attention mechanism. It optimizes the model's attention to key image areas by calculating the weight distribution between input features, and adjusts the model's sensitivity to key parts of the image to generate a focus image.

[0114] The attention mechanism integration submodule uses a multi-head attention mechanism to analyze and weight different feature areas in medical images, allowing the model to focus on multiple areas at the same time when processing images, thereby enhancing the ability to parse key parts of the image. By modeling the dependencies between features, the model can more accurately identify and focus on those image parts that are most critical to diagnosis, such as lesion areas or abnormal structures. Finally, the module outputs an image that highlights the key areas, namely the focus image, which provides a basis for subsequent feature visualization and analysis.

[0115] Multi-head attention mechanism, according to the formula:

[0116] ;

[0117] Calculate weighted output values ​​to highlight key image areas;

[0118] in, is the query matrix, is the key matrix, is the value matrix, The context relevance matrix adds mutual information between input features to improve the sensitivity and accuracy of the model. is the dimension of the key vector, is a small amount, used to avoid division by zero errors, Represents the query matrix With key matrix The transpose of is used for dot product operation, Represents the multi-head self-attention function.

[0119] Execution process:

[0120] First, query the matrix and key matrix Perform a dot product operation to calculate the corresponding score, which maps the relative importance of the input features. Then, add the context relevance matrix ,This matrix is ​​obtained through additional training parameters, combining input features and external information to enhance the model's sensitivity to specific medical image regions such as lesions or abnormal structures. Then, the dot product score and the context score are added and the key vector dimension is passed Adjust, add small amount To prevent division by zero errors, these scores are processed with a softmax function to obtain the final weight distribution. Finally, the weights are applied to the value matrix , generate feature representation of the focus image, weight coefficient The value of is optimized during the training process through back-propagation and gradient descent methods, ensuring that each step of the calculation is aimed at enhancing diagnostic accuracy.

[0121] The feature visualization submodule uses gradient-weighted class activation mapping based on the focus image to highlight the key areas of model decision-making, generates a heat map through a weighted combination of gradient and feature map, and processes the superposition of the heat map and the original image to generate a visual explanation map;

[0122] The feature visualization submodule uses the gradient-weighted class activation mapping (Grad-CAM) technology, which analyzes the convolutional layers in the model to determine which regions play a key role in making diagnostic decisions. By weightedly combining the gradient of the region with the feature map, a heat map showing the focus of the model is generated. This heat map is then superimposed on the original medical image to provide an intuitive visual explanation of how the model understands and analyzes the image, thereby generating a highly interpretable visual explanation map.

[0123] The explanatory feedback loop submodule adjusts the attention mechanism based on the visual explanation graph and the doctor feedback, analyzes the doctor's feedback on the visual output, identifies the inadequacy of the model analysis, and iterates repeatedly to generate an explanatory enhancement model.

[0124] The explanatory feedback loop submodule aims to further improve the model's explanatory power and accuracy by integrating actual feedback from doctors. The feedback provided by doctors focuses on whether the model's explanatory diagram truly reflects the key considerations in diagnosis and important information that the model may overlook. Based on the feedback, the weight and focus of the attention mechanism are adjusted, and the algorithm's parsing logic is optimized. This process is iterative, and through multiple adjustments and optimizations, the performance of the model is gradually improved to ensure that the generated explanatory enhanced model can provide more accurate and explanatory support in actual medical scenarios.

[0125] See also Figure 2 and Fig.10 ,The comprehensive diagnosis module includes the lesion detection submodule, the diagnosis analysis submodule, and the information generation submodule;

[0126] The lesion detection submodule performs initial processing on the input diagnostic data based on the interpretive enhancement model and real-time optimization feedback information, extracts key diagnostic features, and distinguishes lesion from non-lesion areas to generate lesion detection data;

[0127] The lesion detection submodule uses an interpretable enhancement model to process and analyze data received from various medical imaging devices. The model first preprocesses the image, such as denoising and contrast enhancement, to optimize image quality. Then, the model uses its trained algorithm to identify and extract key diagnostic features in the image, such as the shape, size, and edge features of the lesion. Based on the features, the model distinguishes between lesion and non-lesion areas and accurately marks key areas. Through this process, the generated lesion detection data provides a basis for subsequent in-depth analysis.

[0128] The diagnostic analysis submodule analyzes the lesion characteristics based on the lesion detection data, analyzes the relationship between the lesion area characteristics and the known medical diagnostic standards, and generates a preliminary diagnostic analysis record;

[0129] The diagnostic analysis submodule conducts in-depth analysis of the nature and possible medical categories of the lesions based on the detected lesion characteristics, and determines the degree of match between the lesion characteristics and various medical conditions by comparing the medical diagnostic standards in the database. This analysis process involves a comprehensive evaluation of the morphology, biomarkers and other diagnostic indicators of the lesions. Through this information, a detailed preliminary diagnostic analysis record is generated to provide a scientific basis for clinical decision-making.

[0130] The information generation submodule organizes and formats the diagnostic data based on the preliminary diagnostic analysis records, creates text, optimizes text diversity, describes the diagnostic data and the optimal operation plan, and generates lesion diagnosis information;

[0131] The information generation submodule is responsible for converting technical diagnostic records into easy-to-understand medical reports. Through natural language processing technology, the module formats the diagnostic data into structured text while ensuring the accuracy and professionalism of the text. In addition, the module will recommend corresponding treatment strategies or further testing plans based on the diagnostic results to ensure that the information provided is comprehensive, diverse and operationally instructive, and ultimately generate lesion diagnosis information.

[0132] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A medical imaging lesion detection system, characterized in that: The system comprises: The data processing module adjusts the brightness of the image based on the original medical image, optimizes the random noise and image edges in the image, and generates a processed image; The quality control module identifies blurry areas and artifacts in the image based on the processed image, and uses image enhancement technology to refine the image quality and generate a quality-optimized image; The resolution reconstruction module optimizes the image resolution based on the quality optimized image, determines the image authenticity, and iteratively optimizes the network parameters to generate a resolution adjusted image; The data enhancement module adjusts the image based on the resolution, uses a conditional generative adversarial network to simulate the lesion characteristics in the medical image, constructs a synthetic image of the target feature, and expands the original data set in combination with the synthetic image to generate an enhanced data set; The data enhancement module comprises: The conditional GAN ​​development submodule adjusts the image based on the resolution, adopts a conditional generative adversarial network, sets the generator to receive the lesion label as a conditional input, combines the image and lesion information, evaluates the authenticity of the synthesized image and the accuracy of the lesion feature, and generates a lesion feature synthesized image; The synthetic image quality assessment submodule calculates the Laplace variance of the image based on the lesion feature synthetic image to perform accuracy assessment, verifies the matching of the quality of the synthetic image with the analysis requirements, and generates a synthetic image assessment record; The data enhancement strategy optimization submodule rotates, scales and crops the image based on the synthetic image evaluation record to optimize the diversity of the data set, and combines the obtained synthetic image to expand the original data set to generate an enhanced data set; The conditional generation adversarial network is based on the formula: ； Calculate the loss of the discriminator; in, and To adjust the weights, and is the regularization coefficient, For real images, is the lesion label, is a random noise input, and is the probability of the discriminator evaluating real and synthetic images, and The output of the discriminator for real and synthetic images at low-level features; The dynamic learning module adjusts the model parameters based on the enhanced data set and according to the doctor's feedback as a reward signal, optimizes the model error, and generates an optimized recognition model; The real-time feedback module analyzes the collected images in real time based on the optimized recognition model, identifies image quality problems, and feeds back to the operator through the user interface to generate real-time optimization feedback information; The model explanation enhancement module applies a multi-head attention mechanism based on the optimized recognition model, analyzes the image area of ​​the model target, visualizes the area, highlights the model focus, and generates an explanation enhancement model; The model explanation enhancement module includes: The attention mechanism integration submodule is based on the optimized recognition model and adopts a multi-head attention mechanism to optimize the model's attention to key image areas by calculating the weight distribution between input features, and adjusts the model's sensitivity to key parts of the image to generate a focus image; The feature visualization submodule uses gradient weighted class activation mapping based on the focus image to highlight the key areas of model decision, generates a heat map through a weighted combination of gradient and feature map, and processes the superposition of the heat map and the original image to generate a visual explanation map; The explanatory feedback loop submodule adjusts the attention mechanism based on the visual explanation graph and the doctor's feedback, analyzes the doctor's feedback on the visual output, identifies the insufficiency of the model analysis, and iterates repeatedly to generate an explanatory enhancement model; The multi-head attention mechanism is based on the formula: ； Calculate weighted output values ​​to highlight key image areas; in, is the query matrix, is the key matrix, is the value matrix, is the context relevance matrix, is the dimension of the key vector, is a small amount, used to avoid division by zero errors, Represents the query matrix With key matrix The transpose of is used for dot product operation, represents the multi-head self-attention function; The comprehensive diagnosis module analyzes the lesion area based on the explanatory enhancement model and the real-time optimization feedback information, and organizes and formats the diagnosis data to generate lesion diagnosis information.

2. The medical imaging lesion detection system according to claim 1, characterized in that: The processed image includes an adjusted brightness level, an adjusted noise level and an optimized image edge, the quality-optimized image includes a corrected blurred area, eliminated artifacts and refined image details, the enhanced data set includes a synthesized lesion feature image and an expanded number of samples, the optimized recognition model includes an adjusted learning rate, an optimized loss function and a solidified recognition strategy, the explanatory enhancement model includes a visualized area of ​​interest, annotated key feature points and an optimized model resolution, and the lesion diagnosis information includes an analyzed lesion area, a summarized diagnostic logic and formatted diagnostic information content.

3. The medical imaging lesion detection system according to claim 1, characterized in that: The dynamic learning module includes: The deep Q network training submodule performs deep Q network training based on the enhanced data set, including matching the input layer with the image data size, extracting key features in the medical image, and optimizing the training process to generate a preliminary recognition model; The doctor feedback integration submodule integrates the doctor's diagnosis feedback as a reward signal based on the preliminary recognition model, adjusts the model learning target through positive and negative feedback, and matches the model to the medical diagnosis standard to generate a feedback adjustment model; The recognition strategy adjustment submodule adjusts the model based on the feedback, analyzes the doctor's feedback and model performance, identifies the model's weaknesses in identifying the target lesions, optimizes the accuracy of identifying the target lesions, matches it to the doctor's diagnostic ideas, and generates an optimized recognition model.

4. The medical imaging lesion detection system according to claim 1, characterized in that: The real-time feedback module comprises: The real-time image analysis submodule performs image processing based on the optimized recognition model, including adjusting the image size and applying filters to optimize image features, and uses the processed image to perform lesion recognition tasks to generate real-time analysis data; The quality assessment model submodule assesses the quality parameters of the image based on the real-time analysis data, scores the image quality according to the model output, identifies image quality problems according to the score, and generates a quality assessment result; The optimization strategy proposal submodule formulates an image optimization strategy based on the quality assessment results and the identified image quality issues, optimizes the image quality, and feeds back to the operator through the user interface to generate real-time optimization feedback information.

Images