Artificial intelligence medical image auxiliary diagnosis equipment and system thereof

Abstract

The invention relates to the technical field of medical image auxiliary diagnosis, in particular to an artificial intelligence medical image auxiliary diagnosis device and a system thereof, and the system comprises a medical image collection and standardization preprocessing module which is used for obtaining multi-modal medical image data of a target patient, carrying out the denoising, alignment and gray level normalization processing of the image, and obtaining the image data of the target patient; generating a standardized input image; and the feature expression construction module is used for extracting multi-level semantic features in the standardized image based on a deep neural network, forming multi-scale feature expression and outputting tissue structure and focus information. According to the medical image auxiliary diagnosis system, medical image acquisition, feature expression construction, cross-modal knowledge fusion, interpretive auxiliary diagnosis, adaptive optimization update and result output and feedback management modules are integrated into the same system, so that the problem that most of traditional medical image auxiliary diagnosis systems adopt a single module to process images; therefore, the problems of dispersed diagnosis process and low information utilization rate are solved.

Description

Technical Field

[0001] This invention relates to the field of medical image-assisted diagnosis technology, and in particular to an artificial intelligence-assisted medical image diagnosis device and system. Background Technology

[0002] In recent years, with the development of medical imaging technology, multimodal medical imaging such as CT, MRI, and ultrasound has been widely used in clinical diagnosis. To improve diagnostic efficiency and accuracy, artificial intelligence technology has been gradually introduced into the field of medical image-assisted diagnosis. Through deep learning models, feature extraction and lesion identification are performed on images to achieve automated or semi-automated diagnostic analysis.

[0003] Traditional medical image-assisted diagnostic systems mostly use a single module to process images. Due to the lack of collaboration between modules, the diagnostic process is fragmented and the information utilization rate is low. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides an artificial intelligence-assisted medical image diagnosis device and system, which aims to improve the problems of traditional medical image-assisted diagnosis systems, which mostly use a single module to process images. Due to the lack of collaboration between modules, the diagnosis process is fragmented and the information utilization rate is low.

[0005] In a first aspect, the present invention provides the following technical solution: an artificial intelligence-assisted medical image diagnosis system includes: The medical image acquisition and standardized preprocessing module is used to acquire multimodal medical image data of the target patient and perform denoising, alignment and grayscale normalization on the images to generate standardized input images. The feature expression construction module is used to extract multi-level semantic features from the standardized image based on a deep neural network, form a multi-scale feature expression, and output tissue structure and lesion information. The cross-modal knowledge fusion module is used to associate and map the multimodal image features with pathological feature nodes in the medical knowledge graph, and to jointly represent the features and knowledge through a semantic attention mechanism. An interpretable auxiliary diagnostic module is used to generate diagnostic results and corresponding salient region mappings based on the joint representation features using an interpretable model, and output model inference information. The adaptive optimization and update module is used to dynamically adjust the model parameters and knowledge mapping weights based on the consistency of doctor feedback and diagnosis results. The results output and feedback management module is used to store the optimized diagnostic results and visual explanations in the medical database and to use them for subsequent data association processing.

[0006] By adopting the above technical solution, the system integrates medical image acquisition, feature representation construction, cross-modal knowledge fusion, interpretive auxiliary diagnosis, adaptive optimization and updating, and result output and feedback management modules into the same system, thereby realizing end-to-end processing of multimodal medical images. This improves the problem that traditional medical image-assisted diagnosis systems mostly use a single module to process images, and the lack of joint collaboration between modules results in a fragmented diagnostic process and low information utilization.

[0007] Furthermore, the denoising, alignment, and grayscale normalization processing of the image includes the following steps: The original image is denoised, and the denoising process includes Gaussian filtering, median filtering or nonlocal mean filtering; The denoised image is aligned, and the alignment process includes rigid transformation, affine transformation or non-rigid registration. The aligned image is subjected to grayscale normalization processing, which includes linear normalization, Z-score normalization, or histogram matching. The standardized images are output as input for subsequent feature extraction and diagnostic analysis.

[0008] Furthermore, the extraction of multi-level semantic features from the standardized image to form a multi-scale feature representation includes the following steps: Input the standardized image into the convolutional neural network; Low-level, mid-level, and high-level features are extracted through multi-layer convolution. Feature maps at different levels are fused at multiple scales according to spatial resolution to form a multi-scale feature representation. Multi-scale features are integrated through channels or space to form multi-level semantic features; The multi-level semantic features are output as a representation of tissue structure and lesion information.

[0009] Furthermore, the step of associating and mapping the multimodal image features with pathological feature nodes in the medical knowledge graph includes the following steps: Multimodal image features are mapped to a unified vector space through feature embedding; Node vectors are generated by embedding pathological feature nodes in a medical knowledge graph. Calculate the similarity or correlation between image feature vectors and knowledge node vectors to form an image-knowledge mapping matrix.

[0010] Furthermore, the joint representation of features and knowledge through the semantic attention mechanism includes the following steps: Input image feature vectors and knowledge node vectors into the semantic attention mechanism; Calculate attention weights and express the degree of association between image features and knowledge nodes as attention scores; A joint representation vector is generated by weighting and fusing image feature vectors and knowledge node vectors based on attention scores. The joint representation vectors are integrated to form cross-modal joint representation features for subsequent diagnostic analysis.

[0011] Furthermore, the step of generating diagnostic results and corresponding salient region mappings using an interpretable model includes the following steps: Input cross-modal joint representation features into an interpretable model; Diagnostic results are generated from an interpretable model; The contribution of each feature in the joint characterization features to the diagnostic results is calculated using an interpretable method; Feature contributions are mapped to the input image region to generate a salient region map; Output the diagnostic results and the corresponding salient region mapping.

[0012] Furthermore, the dynamic adjustment of model parameters and knowledge mapping weights includes the following steps: Receive doctor feedback and information on the consistency of diagnostic results; Calculate diagnostic consistency indicators; Use optimization algorithms to adjust the parameters of deep neural network models; Adjust the cross-modal knowledge mapping weights based on diagnostic consistency indicators; Output the updated model parameters and knowledge mapping weights.

[0013] Secondly, the present invention provides the following technical solution: an artificial intelligence medical image-assisted diagnostic device, the device comprising a housing and a power management module, the housing being used to house a medical image acquisition and standardization preprocessing module, a feature expression construction module, a cross-modal knowledge fusion module, an interpretative auxiliary diagnostic module, an adaptive optimization and update module, and a result output and feedback management module, and providing physical protection for the modules, the power management module being used to provide power to each module.

[0014] Thirdly, the present invention provides the following technical solution: an artificial intelligence-assisted medical image diagnosis method, the method comprising the following steps: Acquire multimodal medical image data of the target patient, and perform denoising, alignment and grayscale normalization on the images to generate standardized input images; Based on deep neural networks, multi-level semantic features are extracted from the standardized images to form multi-scale feature representations and output tissue structure and lesion information. The multimodal image features are associated and mapped with pathological feature nodes in the medical knowledge graph, and the features and knowledge are jointly represented through a semantic attention mechanism. Based on the joint representation features, the diagnostic results and corresponding salient region mappings are generated using an interpretable model, and the model inference information is output. Based on the consistency between doctor feedback and diagnostic results, the model parameters and knowledge mapping weights are dynamically adjusted. The optimized diagnostic results and visual explanations are stored in a medical database and used for subsequent data association processing.

[0015] The present invention has the following beneficial effects: 1. In this invention, by integrating medical image acquisition, feature representation construction, cross-modal knowledge fusion, interpretive auxiliary diagnosis, adaptive optimization and updating, and result output and feedback management modules into the same system, end-to-end processing of multimodal medical images is achieved. This improves the problem that traditional medical image-assisted diagnosis systems mostly use a single module to process images, and the lack of joint collaboration between modules results in a fragmented diagnostic process and low information utilization.

[0016] 2. In this invention, the original image is denoised, aligned, and normalized to generate a standardized input image. This improves the problem that traditional medical image processing often uses only denoising or does not perform standardization, resulting in unstable feature extraction and large errors in diagnostic results due to inconsistent image quality and spatial scale.

[0017] 3. In this invention, by associating and mapping multimodal image features with pathological feature nodes in a medical knowledge graph, and jointly representing them through a semantic attention mechanism, cross-modal joint representation features are formed. This improves the problem that traditional medical image diagnosis mostly relies solely on image information for analysis, and the diagnosis results lack reference semantic support due to the lack of structured knowledge. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the architecture of an artificial intelligence-assisted medical image diagnosis system proposed in this invention; Figure 2 This is a schematic diagram of the structure of an artificial intelligence-assisted medical image diagnosis device proposed in this invention; Figure 3 This is a flowchart illustrating an artificial intelligence-assisted medical image diagnosis method proposed in this invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: In a first embodiment of the present invention, the present invention provides an artificial intelligence-assisted medical image diagnosis system, such as... Figure 1 As shown, it includes: The medical image acquisition and standardized preprocessing module is used to acquire multimodal medical image data of the target patient and perform denoising, alignment and grayscale normalization on the images to generate standardized input images. Furthermore, the image denoising, alignment, and grayscale normalization processes include the following steps: The original image is denoised, and the denoising process includes Gaussian filtering, median filtering, or nonlocal mean filtering. The denoised image is then aligned, and the alignment process includes rigid transformation, affine transformation, or non-rigid registration. The aligned image is subjected to grayscale normalization, which includes linear normalization, Z-score normalization, or histogram matching. The standardized images are output as input for subsequent feature extraction and diagnostic analysis.

[0021] Specifically, in the medical image acquisition and standardized preprocessing module, multimodal medical image data of the target patient is acquired, including CT, MRI, or ultrasound images. The original images are denoted as... Its dimensions are: ; in and These represent the height and width of the image, respectively. Indicates the number of channels, with pixel values... The images originate from the scan output of medical imaging equipment. To reduce noise interference and improve the stability of subsequent feature extraction, the original images are first denoised. Gaussian filtering, median filtering, or nonlocal mean filtering can be selected. Taking Gaussian filtering as an example, the denoised image... It can be represented as: ; in A two-dimensional Gaussian kernel, Indicates the Gaussian kernel standard deviation. The radius of the convolution window is determined based on the image resolution and noise level. Next, the denoised image undergoes alignment processing to correct for patient positional differences and intermodal spatial inconsistencies, generating an aligned image. Alignment can be achieved using rigid transformation, affine transformation, or non-rigid registration. Taking rigid transformation as an example, its mathematical representation is: ; in For rotation matrix, The translation vector, matrix, and vector are obtained by optimizing the spatial registration loss function, which can be defined as the mean square error or mutual information of the two images. Finally, the aligned image is subjected to grayscale normalization to map pixel values ​​to a uniform range to eliminate brightness differences caused by different scanning devices and parameters, resulting in a standardized image. Linear normalization can be used: ; in and These are the minimum and maximum values ​​of pixels in the image, respectively, or can be normalized using Z-score: ; in and Representing channels The mean and standard deviation of the output image are calculated from the aligned image pixels. As input to the next feature representation construction module, it is used for multi-level semantic feature extraction and multi-modal joint analysis based on deep neural networks, thereby providing unified and quality-controllable image foundation data for subsequent cross-modal knowledge fusion and interpretable assisted diagnosis.

[0022] This reduces image noise, corrects spatial location differences, and unifies grayscale distribution, providing standardized input images with controllable and consistent quality for subsequent feature extraction and diagnostic analysis.

[0023] The feature representation construction module is used to extract multi-level semantic features from standardized images based on deep neural networks, form multi-scale feature representations, and output tissue structure and lesion information. Furthermore, extracting multi-level semantic features from standardized images to form multi-scale feature representations includes the following steps: Input the standardized image into the convolutional neural network; Low-level, mid-level, and high-level features are extracted through multi-layer convolution. Feature maps at different levels are fused at multiple scales according to spatial resolution to form a multi-scale feature representation. Multi-scale features are integrated through channels or space to form multi-level semantic features; Output multi-level semantic features as representations of organizational structure and lesion information.

[0024] Specifically, in the feature representation construction module, standardized images are... Input to a convolutional neural network, image size is ,in Indicates height, Indicates width, The number of channels is represented by pixel values ​​obtained from the preprocessing module. Features at different levels are extracted through multi-layer convolution operations, where the first... Layer features It can be represented as: ; in The input feature map for the previous layer, The convolution kernel weight matrix, For the bias vector, Indicates the convolution operation. For activation functions such as ReLU, low-level features Mainly preserves local texture information, mid-level features Capturing mesoscale structures and high-level features To represent global semantic information, feature maps from different levels are upsampled or downsampled according to spatial resolution and then fused to form a multi-scale feature representation. : ; in Represents the upsampling function. The function represents the downsampling function, and Concat represents concatenation along the channel dimension, followed by multi-scale features. Perform channel integration or spatial integration to generate multi-level semantic features. This can be achieved through convolution or attention mechanisms: ; in This represents the integration function, whose weights are obtained through network training, and the output is... The feature representations of organizational structure and lesion information are used as input to the next cross-modal knowledge fusion module, providing a unified semantic feature foundation for joint representation and interpretable diagnosis.

[0025] This enables the extraction of structural and semantic information at different scales from images, forming a unified multi-level feature representation that provides stable and rich input features for tissue structure analysis and lesion identification.

[0026] The cross-modal knowledge fusion module is used to associate and map multimodal image features with pathological feature nodes in the medical knowledge graph, and to jointly represent features and knowledge through a semantic attention mechanism. Furthermore, the association and mapping of multimodal image features with pathological feature nodes in the medical knowledge graph includes the following steps: Multimodal image features are mapped to a unified vector space through feature embedding; Node vectors are generated by embedding pathological feature nodes in a medical knowledge graph. Calculate the similarity or correlation between image feature vectors and knowledge node vectors to form an image-knowledge mapping matrix.

[0027] Specifically, in the cross-modal knowledge fusion module, the input data includes multi-level image features output by the aforementioned feature representation construction module. and the set of pathological feature nodes in the medical knowledge graph ,in and These represent the height and width of the image feature map, respectively. Indicates the number of channels. To determine the number of pathology nodes in the knowledge graph, node information is constructed from medical literature or clinical data. First, image features are embedded using a feature embedding function. Mapping to a unified vector space yields image feature vectors: ; in Number of pixel positions For the embedding dimension, the embedding weights are obtained through network training, and node embedding is performed on the knowledge graph nodes. Generate node vectors Representing each pathological feature, the embedding function can be obtained based on graph embedding methods such as Graph Convolution Network or Node2Vec. Subsequently, the similarity matrix between the image feature vector and the knowledge node vector is calculated. The cosine similarity formula can be used: ; in Indicates the image number Feature vectors at each position, Indicates the first From the knowledge node vectors, a similarity matrix is ​​obtained, forming an image-knowledge mapping matrix. This is used to represent the degree of association between local image features and knowledge nodes. As the input for the next step of the semantic attention mechanism, it jointly represents features and knowledge, realizes cross-modal information fusion, and provides a feature foundation that combines structured and semantic information for the interpretive auxiliary diagnosis module.

[0028] This enables the association between multimodal image features and pathological information in medical knowledge graphs, achieving joint representation of image features and structured knowledge, and providing semantic support and reference for subsequent interpretable diagnosis.

[0029] Furthermore, the joint representation of features and knowledge through semantic attention mechanisms includes the following steps: Input image feature vectors and knowledge node vectors into the semantic attention mechanism; Calculate attention weights and express the degree of association between image features and knowledge nodes as attention scores; A joint representation vector is generated by weighting and fusing image feature vectors and knowledge node vectors based on attention scores. The joint representation vectors are integrated to form cross-modal joint representation features for subsequent diagnostic analysis.

[0030] Specifically, in the cross-modal knowledge fusion module, the input data includes: Image feature vector and knowledge node vectors ; in Indicates the number of image feature locations. Indicates the number of knowledge nodes. For the embedding dimension, the vectors are obtained through the aforementioned feature embedding and node embedding. After inputting the two sets of vectors into the semantic attention mechanism, the attention weight matrix of image features for each knowledge node is first calculated. The dot product attention formula can be used: ; in Indicates the image number Each location feature vector, Indicates the first The image feature vector and the knowledge node vector are weighted and fused according to the attention weights to generate a joint representation vector: Subsequently, the joint representation vectors at all locations are integrated by channel or spatial methods to obtain cross-modal joint representation features. ;in and Image space size, The number of feature channels after integration is given. The integration operation can be performed using convolutional or fully connected layers. The weights are obtained through network training, and the output is... The joint semantic representation of image features and knowledge graph information is used as input for the next step of the interpretive auxiliary diagnosis module, providing a feature basis that combines structured and semantic information for generating diagnostic results and mapping of salient regions.

[0031] This enables the weighted fusion of image features and medical knowledge nodes to form a cross-modal joint representation, realizing the semantic combination of images and structured knowledge, and providing a comprehensive feature foundation for interpretable diagnosis.

[0032] The interpretable auxiliary diagnosis module is used to generate diagnostic results and corresponding salient region mappings based on joint representation features using an interpretable model, and output model inference information. Furthermore, generating diagnostic results and corresponding salient region mappings using an interpretable model includes the following steps: Input cross-modal joint representation features into an interpretable model; Diagnostic results are generated from an interpretable model; The contribution of each feature in the joint characterization features to the diagnostic results is calculated using an interpretable method; Feature contributions are mapped to the input image region to generate a salient region map; Output the diagnostic results and the corresponding salient region mapping.

[0033] Specifically, in the interpretive auxiliary diagnosis module, the input data consists of cross-modal joint representation features: ; in and Indicates the image space size. The number of feature channels is output by the aforementioned cross-modal knowledge fusion module. Input interpretable model The model can use linear interpretable models, weighted eigenmaps, or gradient basis methods to generate diagnostic predictions. : ; in , The number of diagnostic categories is determined by the prediction results obtained through model training. Subsequently, an interpretable method is used to calculate the contribution of each joint representation feature to the diagnostic result; a typical method is gradient-weighted feature contribution. ; in Indicates position aisle eigenvalues, The gradient is calculated using backpropagation to obtain the predicted value for the corresponding category, taking into account the contribution of each channel. Summation or weighted integration along the channel dimension generates a salient region mapping. : ; in The channel weights can be obtained through training or by assigning equal values ​​to the output. For the diagnostic results, The corresponding saliency region mappings are used to provide specific lesion or disease classification information, and the latter to visualize the diagnostic basis. The generated diagnostic results and saliency mappings can be used as inputs for doctor review and subsequent adaptive optimization update modules to dynamically adjust model parameters and knowledge mapping weights, thereby improving the reliability and interpretability of the diagnosis.

[0034] This enables the generation of specific diagnostic results based on joint features, while providing salient region mapping, thus achieving interpretability and visualization of diagnostic results and providing a reference for clinical judgment.

[0035] The adaptive optimization and update module is used to dynamically adjust the model parameters and knowledge mapping weights based on the consistency of doctor feedback and diagnosis results. Furthermore, dynamically adjusting the model parameters and knowledge mapping weights includes the following steps: Receive doctor feedback and information on the consistency of diagnostic results; Calculate diagnostic consistency indicators; Use optimization algorithms to adjust the parameters of deep neural network models; Adjust the cross-modal knowledge mapping weights based on diagnostic consistency indicators; Output the updated model parameters and knowledge mapping weights.

[0036] Specifically, in the adaptive optimization and update module, the input data includes: Doctor feedback information Consistency information with diagnostic results ; in This indicates the number of diagnostic categories. Feedback information is obtained through doctors' annotations or ratings of previous diagnostic results. Consistency information is generated by comparing the previous diagnostic result with the current model output. First, the diagnostic consistency index is calculated. : ; in The current output class probability of the model. For doctor feedback values, For difference measures such as mean squared error or cross entropy, according to Update deep neural network parameters using optimization algorithms such as gradient descent or Adam. : ; in For learning rate, To combine the loss function of diagnostic consistency and classification error, and to adjust the cross-modal knowledge mapping weights based on the consistency index. : ; in Image-knowledge mapping matrix, Given the weight update rate, output the updated weights. and This technology is used for the next round of joint representation feature generation and interpretive assisted diagnosis, enabling dynamic adaptive optimization of model parameters and knowledge mapping. This allows for gradual adjustment of the model in multiple feedback loops, improving the matching degree between diagnostic results and physician opinions, and providing a reliable update mechanism for clinical assistance.

[0037] This allows for dynamic adjustment of model parameters and knowledge mapping weights based on doctor feedback and diagnostic consistency, enabling adaptive optimization of the model and improving the consistency and reliability of diagnostic results with clinical judgment.

[0038] The results output and feedback management module is used to store the optimized diagnostic results and visual explanations in the medical database and to use them for subsequent data association processing.

[0039] Specifically, in the results output and feedback management module, the input data includes: Optimized diagnostic results and corresponding visual explanation information ; in Indicates the number of diagnostic categories. and The image spatial dimensions are generated by the interpretative diagnostic module, which first... and Structured encapsulation is performed to form record entries: ; in For diagnosis timestamps, A unique identifier for each patient is written into the medical database via a database interface. : Subsequently, the stored records are used for further data association processing, which can form an association matrix by searching for similar cases or conducting statistical analysis. : ; in For the number of records in the database, The function used is a feature similarity function, such as cosine similarity or Euclidean distance normalization. The result is calculated based on the integration of diagnostic result vectors and salient region mapping features. The output is a structured diagnostic record stored in the database. and the correlation matrix used for analysis These results are used to support the feedback loop of physician review, statistical analysis, case retrieval, and subsequent adaptive optimization and update modules, providing a data foundation and historical reference for clinical decision-making.

[0040] This allows for the systematic storage of optimized diagnostic results and visual explanations, supporting subsequent case association analysis and data reuse, and providing a data foundation for clinical decision-making and adaptive model optimization.

[0041] like Figure 2 An artificial intelligence-assisted medical image diagnosis device is disclosed. The device includes a housing and a power management module. The housing is used to house a medical image acquisition and standardized preprocessing module, a feature expression construction module, a cross-modal knowledge fusion module, an interpretive auxiliary diagnosis module, an adaptive optimization and update module, and a result output and feedback management module, and provides physical protection for the modules. The power management module is used to provide power to each module.

[0042] Specifically, the outer shell houses and secures the medical image acquisition and standardized preprocessing module, the feature representation construction module, the cross-modal knowledge fusion module, the interpretative auxiliary diagnosis module, the adaptive optimization and update module, and the result output and feedback management module. Physical support and protective structures ensure the stability of each module and prevent interference between them. The input data consists of the patient's multimodal medical images. and doctor's feedback ,in and Image height and width, respectively For the number of channels, To determine the number of feedback categories, images are acquired by medical imaging equipment, and doctor feedback is generated through manual annotation or scoring. The power management module controls the input voltage. Adjust to the required voltage for each module and provide current Ensure continuous operation: ; in and The electrical characteristic parameters of each module are determined by the module specifications. After receiving input data, each functional module operates according to its own processing flow. For example, the standardized preprocessing module performs image preprocessing to generate standardized images. The feature representation construction module extracts multi-level features. The cross-modal knowledge fusion module generates cross-modal joint representations. The interpretive auxiliary diagnostic module outputs diagnostic results. and salient region mapping The adaptive optimization and update module updates based on feedback. and consistency indicators Adjusting model parameters And knowledge mapping weight The results output and feedback management module will output the results. and Stored in database This data is then used for subsequent association analysis, and the output is a structured diagnostic record. It is used to support clinical decision-making, case retrieval, and model adaptive optimization loops, thereby realizing the overall automated diagnostic function and continuous performance optimization of the device.

[0043] This allows for the centralized housing of various functional modules and provides stable power and physical protection, enabling continuous and efficient AI-powered medical image processing and assisted diagnosis, and supporting clinical data management and analysis.

[0044] Example 2: Early breast cancer screening faces challenges such as high image noise, difficulty in uniformly analyzing images of different modalities, strong subjectivity in physician diagnoses, and inaccurate lesion localization. To address these issues, this invention employs an artificial intelligence-assisted medical image diagnosis method, the process of which is as follows: Figure 3 As shown. The specific implementation process of this method is as follows: First, acquire multimodal image data of the target patient’s mammogram, ultrasound and MRI, and perform noise reduction, alignment and grayscale normalization on the images to generate standardized input images, thereby improving image quality and ensuring the comparability of different modal images in space and grayscale. Secondly, based on deep neural networks, multi-level semantic features are extracted from standardized images to form multi-scale feature representations, outputting tissue structure and lesion information in order to clearly present the internal tissue structure of the breast and potential lesion areas; Next, the multimodal image features are associated and mapped with the pathological feature nodes in the breast disease knowledge graph, and the features and knowledge are jointly represented through a semantic attention mechanism to achieve the fusion of image features and medical knowledge, thereby improving the completeness and understandability of diagnostic information. Subsequently, based on the joint representation features, the interpretable model is used to generate diagnostic results and corresponding salient region mappings, outputting model inference information so that doctors can intuitively understand the basis of the model's judgment and make auxiliary decisions. Then, based on the consistency between the doctor's feedback and the diagnosis results, the model parameters and knowledge mapping weights are dynamically adjusted to improve the model's adaptability and diagnostic consistency. Finally, the optimized diagnostic results and visual explanations are stored in a medical database and used for subsequent data association processing, enabling case data accumulation, trend analysis, and clinical decision support. The method provided by this invention can effectively address the problems of high image noise, difficulty in cross-modal information integration, and strong diagnostic subjectivity in early breast cancer screening scenarios.

[0045] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An artificial intelligence medical image assisted diagnosis system, characterized in that, The method comprises the following steps: a medical image acquisition and standardized preprocessing module is used to acquire multi-modal medical image data of a target patient, and to perform denoising, alignment and grayscale normalization processing on the images to generate standardized input images; a feature expression construction module is used to extract multi-level semantic features in the standardized images based on a deep neural network, form multi-scale feature expressions, and output tissue structure and lesion information; a cross-modal knowledge fusion module is used to associate and map the multi-modal image features with pathological feature nodes in a medical knowledge graph, and to jointly represent the features and knowledge through a semantic attention mechanism; an explanatory auxiliary diagnosis module is used to generate diagnosis results and corresponding saliency region mappings using an interpretable model based on the joint representation features, and to output model reasoning information; an adaptive optimization and update module is used to dynamically adjust model parameters and knowledge mapping weights according to doctor feedback and diagnosis result consistency; a result output and feedback management module is used to store the optimized diagnosis results and visualized explanation information in a medical database, and to perform subsequent data association processing. 2.The artificial intelligence medical image auxiliary diagnosis system of claim 1, wherein, The denoising, alignment and grayscale normalization processing of the images comprises the following steps: performing denoising processing on the original images, wherein the denoising processing comprises Gaussian filtering, median filtering or non-local mean filtering; performing alignment processing on the denoised images, wherein the alignment processing comprises rigid transformation, affine transformation or non-rigid registration; performing grayscale normalization processing on the aligned images, wherein the grayscale normalization processing comprises linear normalization, Z-score standardization or histogram matching; outputting the standardized images as inputs for subsequent feature extraction and diagnosis analysis. 3.The artificial intelligence medical image auxiliary diagnosis system of claim 1, wherein, The multi-level semantic feature extraction in the standardized images and the formation of multi-scale feature expressions comprise the following steps: inputting the standardized images into a convolutional neural network; extracting low-level features, middle-level features and high-level features through multiple layers of convolution; performing multi-scale fusion of feature maps at different levels according to spatial resolution to form multi-scale feature expressions; integrating the multi-scale features in the channel or space to form multi-level semantic features; outputting the multi-level semantic features as representations of tissue structure and lesion information. 4.The artificial intelligence medical image auxiliary diagnosis system of claim 1, wherein, The association and mapping of the multi-modal image features with pathological feature nodes in the medical knowledge graph comprise the following steps: mapping the multi-modal image features to a unified vector space through feature embedding; performing node embedding on the pathological feature nodes in the medical knowledge graph to generate node vectors; calculating the similarity or correlation between the image feature vectors and the knowledge node vectors to form an image-knowledge mapping matrix. 5.The artificial intelligence medical image auxiliary diagnosis system of claim 1, wherein, The joint representation of features and knowledge through the semantic attention mechanism comprises the following steps: inputting the image feature vectors and the knowledge node vectors into the semantic attention mechanism; calculating attention weights and representing the degree of association of the image features with the knowledge nodes as attention scores; weighting and fusing the image feature vectors and the knowledge node vectors according to the attention scores to generate joint representation vectors; integrating the joint representation vectors to form cross-modal joint representation features for subsequent diagnosis analysis. 6.The artificial intelligence medical image auxiliary diagnosis system of claim 1, wherein, The method for generating the diagnosis result and the corresponding significant region mapping by using the explainable model comprises the following steps: inputting the cross-modal joint representation features into the explainable model; generating the diagnosis result by the explainable model; calculating the contribution of each feature in the joint representation features to the diagnosis result according to the explainable method; mapping the feature contribution to the input image region to generate the significant region mapping; outputting the diagnosis result and the corresponding significant region mapping. 7.The artificial intelligence medical image auxiliary diagnosis system of claim 1, wherein, The method for dynamically adjusting the model parameters and the knowledge mapping weight comprises the following steps: receiving the doctor feedback and the diagnosis result consistency information; calculating the diagnosis consistency index; adjusting the deep neural network model parameters by using the optimization algorithm; adjusting the cross-modal knowledge mapping weight according to the diagnosis consistency index; outputting the updated model parameters and the knowledge mapping weight.

8. An artificial intelligence medical image assisted diagnosis device, characterized by, The artificial intelligence medical image auxiliary diagnosis system according to any one of claims 1-7, the device comprises a shell and a power management module, the shell is used for accommodating a medical image acquisition and standardized pretreatment module, a feature expression construction module, a cross-modal knowledge fusion module, an explanatory auxiliary diagnosis module, a self-adaptive optimization update module and a result output and feedback management module, and providing physical protection for the modules, and the power management module is used for providing power supply for each module.

9. An artificial intelligence medical image auxiliary diagnosis method, characterized in that, The artificial intelligence medical image auxiliary diagnosis system according to any one of claims 1-7, the method comprises the following steps: acquiring the multi-modal medical image data of the target patient, and performing denoising, alignment and gray scale normalization processing on the image to generate standardized input images; extracting multi-level semantic features in the standardized images based on a deep neural network, forming multi-scale feature expression, and outputting tissue structure and lesion information; associating the multi-modal image features with the pathological feature nodes in the medical knowledge graph, and performing joint representation of the features and the knowledge through a semantic attention mechanism; generating the diagnosis result and the corresponding significant region mapping by using the explainable model based on the joint representation features, and outputting the model reasoning information; dynamically adjusting the model parameters and the knowledge mapping weight according to the doctor feedback and the diagnosis result consistency; storing the optimized diagnosis result and the visual explanation information in the medical database, and using the information for subsequent data association processing.

Images