Medical image intelligent auxiliary diagnosis health management system based on deep learning

Abstract

The invention provides a medical image intelligent auxiliary diagnosis health management system based on deep learning, and relates to the field of medical information.The method comprises the steps that firstly, image data are preprocessed to eliminate equipment noise and physiological artifact interference, and multi-scale space-semantic features of cross-time-phase images are extracted; the method comprises the following steps: capturing a lesion evolution trajectory by using feature map differential operation, performing pathological feature enhancement and redundancy suppression on difference features through spatial-semantic joint enhancement, and finally converting time sequence difference features into a color gradient map according with cognitive habits of radiologists by means of a generation mechanism. Therefore, misjudgment caused by motion artifacts and physiologic changes can be overcome, the spatial topological structure of a focus area is kept, and meanwhile, the distinguishability of tiny pathological changes is enhanced, so that tiny but key pathological feature expression is enhanced, and the original image anatomical sites are visually mapped.

Description

Technical Field

[0001] This application relates to the field of medical information, and more particularly, in the embodiments of this application, it relates to an intelligent auxiliary diagnosis and health management system for medical images based on deep learning. Background Art

[0002] Medical imaging (medical images) refers to the technology and processing process of obtaining images of internal tissues of the human body or a part thereof in a non-invasive manner, and is widely used in medical diagnosis and medical research. It includes a variety of imaging techniques, such as X-ray, ultrasound, CT scan, magnetic resonance imaging (MRI), and radionuclide imaging. These techniques can generate detailed images of the internal structure of the human body to help doctors identify diseases, injuries, and other physiological abnormalities.

[0003] Patent CN118692633A proposes an artificial intelligence-assisted system for medical image diagnosis. In the historical comparison unit, first, historical image data such as the patient's MRI and CT are collected and a series of preprocessing steps are performed on them. Next, in order to quantify the similarity or difference between the current and historical images, the system uses comparison strategies such as the structural similarity index (SSI) or change detection algorithms. These methods usually rely on image processing libraries in MATLAB or Python to implement the calculation process, providing a basis for medical diagnosis and treatment planning.

[0004] Although simple metrics such as the structural similarity index (SSI) have a certain role in quantifying the similarity or difference between images, they are very sensitive to noise and artifacts in the images and are prone to misjudgment. For example, in the presence of slight motion artifacts, the SSI may overestimate the difference between two images. In addition, these traditional algorithms mainly rely on pixel-level or local block-level information for comparison and are difficult to capture complex pattern changes or small but important pathological features, thus affecting the accuracy of processing.

[0005] Therefore, an optimized intelligent auxiliary diagnosis and health management solution for medical images based on deep learning is desired. Summary of the Invention

[0006] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a medical image intelligent auxiliary diagnosis health management system based on deep learning, which first eliminates equipment noise and physiological artifact interference by preprocessing the image data, and extracts multi-scale spatial-semantic features of cross-phase images, and uses feature map difference operations to capture the lesion evolution trajectory, and then uses spatial-semantic joint enhancement to enhance the pathological features and suppress redundancy of the difference features, and finally uses the generation mechanism to convert the temporal difference features into a color gradient map that conforms to the cognitive habits of radiologists. In this way, it is possible to overcome the misjudgment caused by motion artifacts and physiological changes, and enhance the recognizability of subtle pathological changes while maintaining the spatial topological structure of the lesion area, so that it not only enhances the expression of subtle but critical pathological features, but also intuitively maps to the original image anatomical site.

[0007] According to one aspect of the present application, a medical image intelligent assisted diagnosis health management system based on deep learning is provided, which includes:

[0008] A medical imaging data acquisition module, used to acquire current medical imaging data and historical medical imaging data of a patient object;

[0009] A preprocessing module, used for performing image preprocessing on the current medical image data and the historical medical image data to obtain preprocessed current medical image data and preprocessed historical medical image data;

[0010] A difference feature calculation module, used for performing a difference calculation based on image features on the preprocessed current medical image data and the preprocessed historical medical image data to obtain a medical image difference feature map;

[0011] A difference enhancement mapping module is used to perform difference feature enhancement and color mapping on the medical image difference feature map to obtain a visualized difference heat map, wherein the difference enhancement mapping module includes: an enhancement unit, used to perform image enhancement based on feature space-semantic dual-dimensional constraints on the medical image difference feature map to obtain an enhanced medical image difference feature map; and a generation unit, used to map the enhanced medical image difference feature map to obtain the visualized difference heat map.

[0012] Compared with the prior art, the present application provides a medical image intelligent auxiliary diagnosis health management system based on deep learning. It first eliminates equipment noise and physiological artifacts by preprocessing the image data, extracts multi-scale spatial-semantic features of cross-phase images, and uses feature map differential operations to capture the lesion evolution trajectory. Then, the pathological features are enhanced and redundant suppression is performed on the difference features through spatial-semantic joint enhancement, and finally, the temporal difference features are converted into color gradient maps that conform to the cognitive habits of radiologists with the help of a generation mechanism. In this way, it is possible to overcome the misjudgment caused by motion artifacts and physiological changes, and enhance the recognizability of subtle pathological changes while maintaining the spatial topological structure of the lesion area, so that the expression of subtle but critical pathological features is enhanced and intuitively mapped to the original image anatomical site. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0014] Figure 1 This is a system block diagram of a medical image intelligent assisted diagnosis health management system based on deep learning according to an embodiment of the present application.

[0015] Figure 2 This is a data flow diagram of a medical image intelligent assisted diagnosis health management system based on deep learning according to an embodiment of the present application.

[0016] Figure 3 This is a block diagram of a difference enhancement mapping module in a deep learning-based medical image intelligent assisted diagnosis health management system according to an embodiment of the present application.

[0017] Figure 4 This is a block diagram of an enhancement unit in a medical image intelligent assisted diagnosis health management system based on deep learning according to an embodiment of the present application.

[0018] Figure 5 This is a block diagram of an explicit modeling modulation subunit in a deep learning-based medical image intelligent assisted diagnosis health management system according to an embodiment of the present application. DETAILED DESCRIPTION

[0019] Various exemplary embodiments, features and aspects of the present application will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0020] As used herein, the term "exemplary" means "serving as an example, instance, or illustration". Any embodiment described as "exemplary" herein should not be construed as superior to or better than other embodiments.

[0021] In addition, for a better illustration of the present application, numerous specific details are set forth in the following detailed description. It should be understood by those skilled in the art that the present application can be practiced without some of these specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art are not described in detail so as to highlight the gist of the present application.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present application, "a plurality" means two or more unless specifically defined otherwise.

[0023] Medical imaging technologies, such as X-ray, ultrasound, CT scan, MRI, and nuclear imaging, support medical diagnosis and research by generating detailed images of the internal structures of the human body. Patent CN118692633A introduces an artificial intelligence-assisted system for medical image diagnosis. This system collects and preprocesses historical image data of patients, such as MRI and CT, in a historical comparison unit, and uses the structural similarity index (SSI) or change detection algorithms to quantify the similarity or difference between the current and historical images to support diagnosis and treatment planning. However, these methods are sensitive to noise and artifacts and mainly rely on pixel-level information for comparison, making it difficult to effectively capture complex pattern changes or small but critical pathological features, which may lead to misjudgment and affect the diagnostic accuracy. Therefore, although metrics such as SSI are helpful for preliminary analysis, there is still room for improvement in terms of accuracy.

[0024] Based on this, the present application proposes a medical image intelligent auxiliary diagnosis health management system based on deep learning, and constructs a visualization method that integrates deep feature comparison and dynamic enhancement. Specifically, it uses deep learning-based image recognition and analysis technology to first eliminate equipment noise and physiological artifact interference by preprocessing the image data. After that, it extracts multi-scale spatial-semantic features of cross-phase images, uses feature map differential operations to capture the evolution trajectory of lesions, and then uses spatial-semantic joint enhancement to enhance pathological features and suppress redundancy of difference features. Finally, with the help of a generation mechanism, the temporal difference features are converted into a color gradient map that conforms to the cognitive habits of radiologists. This solution replaces traditional pixel-level comparison with deep learning feature representation, effectively overcomes the misjudgment caused by motion artifacts and physiological changes, and enhances the recognizability of subtle pathological changes while maintaining the spatial topological structure of the lesion area, so that the dynamic comparison results of follow-up images can not only enhance the expression of subtle but critical pathological features, but also be intuitively mapped to the original image anatomical site, providing clinical decision-making support with both quantitative analysis and visual interpretability.

[0025] In response to the above technical problems, this application proposes a medical image intelligent assisted diagnosis health management system based on deep learning. Figure 1 This is a system block diagram of a medical image intelligent assisted diagnosis health management system based on deep learning according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of a medical image intelligent assisted diagnosis health management system based on deep learning according to an embodiment of the present application. Figure 1 and Figure 2 As shown, according to the deep learning-based medical image intelligent assisted diagnosis health management system 100 according to the embodiment of the present application, it includes: a medical image data acquisition module 110, which is used to acquire the current medical image data and historical medical image data of the patient object; a preprocessing module 120, which is used to perform image preprocessing on the current medical image data and the historical medical image data to obtain the preprocessed current medical image data and the preprocessed historical medical image data; a difference feature calculation module 130, which is used to perform image feature-based difference calculation on the preprocessed current medical image data and the preprocessed historical medical image data to obtain a medical image difference feature map; a difference enhancement mapping module 140, which is used to perform difference feature enhancement and color mapping on the medical image difference feature map to obtain a visual difference heat map.

[0026] In the above-mentioned intelligent auxiliary diagnosis and health management system 100 for medical images based on deep learning, the medical image data acquisition module 110 is used to acquire the current medical image data and historical medical image data of a patient object. Specifically, the medical image data can be CT, magnetic resonance imaging (MRI), etc. It should be understood that considering the comparative analysis of image data at different times, the development trajectory and changes of lesions can be effectively identified, thereby providing strong support for disease diagnosis, treatment planning, and efficacy evaluation. Therefore, the technical solution of this application acquires the current medical image data and historical medical image data of a patient object. Specifically, in the embodiment of this application, the historical medical image data specifically refers to the data collected immediately before the current medical image data. This means that this application focuses on the image changes at consecutive time points to facilitate the capture of possible pathological feature evolutions in the short term. For example, in the scenario of tumor monitoring, regular CT or MRI examinations can help doctors understand the change trends of structure size, shape, and internal structure, which is of great significance for judging the treatment effect and adjusting the treatment plan. Specifically, in order to acquire these image data, medical institutions usually use advanced imaging technologies, such as computed tomography (CT), magnetic resonance imaging (MRI), etc., to generate detailed images of the internal structure of the human body. Among them, CT scans can quickly provide high-resolution anatomical information and are particularly suitable for observing bone details and lung conditions; while MRI is known for its excellent soft tissue contrast and is widely used in the fields of the nervous system, musculoskeletal system, etc.

[0027] In the above-mentioned intelligent auxiliary diagnosis and health management system 100 for medical images based on deep learning, the preprocessing module 120 is used to perform image preprocessing on the current medical image data and historical medical image data to obtain preprocessed current medical image data and preprocessed historical medical image data. In the embodiments of the present application, the image preprocessing includes denoising filtering, artifact correction, contrast enhancement, and normalization processing. It should be understood that considering that different imaging devices or scanning parameters may be used for multiple examinations of patients, the original images often have problems such as device noise, motion artifacts, and gray-scale distribution offset. For example, the stripe artifacts caused by metal implants in CT images will distort the local anatomical structure, the blurring effect caused by the patient's respiratory movement during MRI scanning may cover up minute features, and the device calibration differences between different medical institutions will cause significant offsets in the contrast of the same tissue in the images. If these interference factors are not eliminated, it will not only cause false alarm differences in traditional SSI and other indicators, but more importantly, it will cause the deep neural network to confuse the noise pattern with the true structural features during feature extraction, thereby affecting the accuracy of the time-series evolution analysis of the entire system. Based on this, in the technical solution of the present application, image preprocessing is performed on the current medical image data and historical medical image data to obtain preprocessed current medical image data and preprocessed historical medical image data. In particular, the image preprocessing includes denoising filtering, artifact correction, contrast enhancement, and normalization processing. Specifically, the denoising filtering adopts an adaptive non-local mean algorithm, which dynamically adjusts the filtering intensity according to the local structure characteristics of the image, and retains the fine anatomical structure while eliminating high-frequency noise; the artifact correction module integrates physical model-driven and deep learning compensation technologies to perform directional repair on specific interference sources such as metal implant artifacts (CT) and motion artifacts (MRI); the contrast enhancement decomposes the reflection component and illumination component of the image through a multi-scale Retinex algorithm to enhance the inherent contrast between tissues without introducing excessive sharpening artifacts; the normalization processing implements non-linear registration based on anatomical landmark points and gray-scale histogram matching to map images at different times to a unified spatial coordinate and gray-scale distribution space. In this way, not only the systematic deviation caused by device heterogeneity is eliminated, but more importantly, a standardized benchmark for cross-temporal images is established to ensure that the subsequent feature extraction network can analyze the anatomical structure at the same scale.

[0028] In the above-mentioned intelligent assisted diagnosis and health management system 100 for medical images based on deep learning, the differential feature calculation module 130 is used to perform differential calculation based on image features on the preprocessed current medical image data and the preprocessed historical medical image data to obtain a medical image differential feature map, including: using an image feature extractor based on serpentine convolution to perform feature extraction on the preprocessed current medical image data and the preprocessed historical medical image data respectively to obtain a current medical image data feature map and a historical medical image data feature map; calculating the position-wise difference between the current medical image data feature map and the historical medical image data feature map to obtain the medical image differential feature map.

[0029] Specifically, an image feature extractor based on serpentine convolution is used to extract features from the preprocessed current medical image data and the preprocessed historical medical image data respectively to obtain a current medical image data feature map and a historical medical image data feature map. It should be understood that considering that lesions in medical images, such as vascular bifurcations, edges of burr-like / lobulated structures, etc., often present irregular geometric shapes, and the subtle progression of chronic lesions often manifests as local textures. The insufficient adaptability of traditional convolutional neural networks to complex structures has become a key bottleneck restricting the accuracy of feature representation. Specifically, traditional rectangular convolution kernels are difficult to fit the organ deformation or the tortuous contour of lesions under a fixed receptive field, and it is easy to lose the continuous features of lesion edges or misidentify the anatomical structure deformation as a differential signal during feature extraction. For example, when analyzing the evolution of irregular structures such as bronchiectasis in lung CT images, conventional convolution operations may not be able to fit the irregular expansion shape of the lumen due to the fixed kernel shape, resulting in missed detection of the true structural differences between adjacent temporal images. Therefore, in the technical solution of this application, an image feature extractor based on serpentine convolution is used to extract features from the preprocessed current medical image data and the preprocessed historical medical image data respectively to obtain a current medical image data feature map and a historical medical image data feature map. Specifically, serpentine convolution is a special convolutional neural network technology designed to improve the flexibility and adaptability of convolution operations by mimicking the path of a snake. This technology is particularly suitable for processing data with complex shapes, especially slender and tortuous structures, such as tubular structures like blood vessels and roads. Considering the particularity of medical image data lies in its complex anatomical structures, organs and lesions often have irregular shapes, forms and position changes, and traditional convolution kernels may be difficult to accurately capture this information. Serpentine convolution allows the shape and size of the convolution kernel to be dynamically adjusted according to the characteristics of the input data. This means that the convolution kernel can "deform" to better match the shape of the target object, thereby improving the accuracy of feature extraction. Specifically, by dynamically adjusting the shape and size of the convolution kernel, serpentine convolution can more effectively capture the local features of the target object, especially when dealing with objects with irregular shapes or complex internal structures. This dynamic adjustment not only improves the ability to capture features but also enhances the adaptability of the model to different shapes and structures. In a specific implementation, the network generates a pixel-level offset vector field through end-to-end learning, driving the convolution kernel to adaptively deform along the morphological contour of the target anatomical structure. For example, when processing lung CT images, the serpentine convolution kernel can fit the burr-like edge of a subpleural nodule and capture the radial texture features of the radial texture expansion through continuous deformation; when analyzing cerebral vascular MRI, the convolution path can be adjusted along the anatomical course of the Willis circle to accurately extract the abnormal deformation features of the arterial wall. This deformation ability is not only reflected in single-frame image processing but also ensures the spatial correspondence of historical and current image feature maps through the deformation consistency constraint of cross-temporal images, establishing a reliable comparison benchmark for subsequent difference detection.

[0030] Specifically, calculate the position-wise difference between the current medical image data feature map and the historical medical image data feature map to obtain the medical image difference feature map. It should be understood that in order to highlight the differences between the current image and the historical image, the differences between the two need to be calculated. However, traditional methods cannot distinguish physiological changes (such as natural displacement of organs) from pathological changes (abnormal anatomical structure changes) through pixel-level or local block-level difference calculations. For example, in the follow-up comparison of lung CT, traditional algorithms may misjudge the deformation of lung lobes caused by respiratory phase differences as regional morphological dilation, while ignoring the subtle heterogeneity changes in the internal density of ground-glass nodules. Based on this, the present application calculates the position-wise difference between the current medical image data feature map and the historical medical image data feature map to obtain the medical image difference feature map. In this way, by comparing the feature maps of the current and historical image data, such changes can be accurately captured. Since the feature maps contain rich structural, texture and other information in the images, calculating the differences between them can present the evolution of image features in the form of feature differences, rather than just an intuitive comparison based on the original images. Specifically, calculating the position-wise difference between the current medical image data feature map and the historical medical image data feature map involves comparing the feature values at corresponding positions pixel by pixel and calculating the differences between these values. This means that for each pair of corresponding pixels (i.e., pixels located at the same spatial position), the system will perform a subtraction operation, subtracting the pixel value in the historical medical image data feature map from the pixel value in the current medical image data feature map at the same position. The result generated in this process is a new numerical value, which represents the relative change amount that has occurred at that position over time. The result of this series of operations is a brand-new feature map, namely the medical image difference feature map, which contains the change information of the feature values at each position. These change information not only reflect the natural evolution of physiology, but may also indicate the abnormal development of pathology, such as the growth or atrophy of tumors, the spread of inflammatory regions, etc.

[0031] Figure 3 It is a block diagram of the difference enhancement mapping module in the intelligent auxiliary diagnosis and health management system for medical images based on deep learning according to an embodiment of the present application. As Figure 3 shown, the difference enhancement mapping module 140 is used to perform difference feature enhancement and color mapping on the medical image difference feature map to obtain a visualized difference heat map, including: an enhancement unit 141 for performing image enhancement based on feature space-semantic two-dimensional constraints on the medical image difference feature map to obtain an enhanced medical image difference feature map; a generation unit 142 for mapping the enhanced medical image difference feature map to obtain the visualized difference heat map.

[0032] Figure 4The block diagram of the enhancement unit in the intelligent auxiliary diagnosis health management system for medical images based on deep learning according to an embodiment of the present application. As Figure 4As shown in the figure, in the embodiment of the present application, the enhancement unit 141 includes: a medical image difference channel feature extraction subunit 1411, configured to extract the channel feature vector at the (i,j) pixel position from the medical image difference feature map as the medical image difference channel feature vector to be enhanced; a medical image difference reference feature random scanning subunit 1412, configured to perform n random scans on the medical image difference feature map to obtain n medical image difference channel feature vectors as a sparse set of medical image difference reference feature vectors; a dominant modeling modulation subunit 1413, configured to perform dominant modeling modulation on the medical image difference channel feature vector to be enhanced and the sparse set of medical image difference reference feature vectors to obtain a medical image difference enhancement component implicit coding vector; a medical image feature fusion subunit 1414, configured to fuse the medical image difference enhancement component implicit coding vector and the medical image difference channel feature vector to be enhanced to obtain an enhanced medical image difference channel feature vector, where the enhanced medical image difference channel feature vector is the channel feature vector at the pixel position (i,j) of the enhanced medical image difference feature map. Considering that the original difference feature map often contains multi-source interference and abnormal structure signals. Due to the spatio-temporal heterogeneity of medical image acquisition (such as equipment parameter drift, patient position difference) and the physiological changes of biological tissues themselves (such as organ periodic movement), simple feature difference operations will introduce a large amount of noise in the difference map. For example, in liver CT follow-up, the diaphragm position shift caused by the difference in respiratory depth may be misidentified as a space-occupying lesion, while the microvascular infiltration characteristics of early hepatocellular carcinoma are submerged in background noise due to weak contrast. At the same time, traditional enhancement methods only rely on a single dimension (such as spatial continuity or semantic salience) for feature screening, and it is difficult to balance the relationship between local anatomical structure changes and global pathological pattern evolution, resulting in the enhanced difference map may over-enhance anatomical displacement artifacts but weaken the invasive growth characteristics of malignant tumors. Based on this, the present application performs image enhancement based on feature space-semantic two-dimensional constraints on the medical image difference feature map to obtain an enhanced medical image difference feature map. That is to say, first, a sparse reference feature library is established in the medical image difference feature map through a random scanning strategy, which not only pays attention to the significant changes in the high-difference response area but also captures the associated changes in the surrounding tissues. The ability of the Poincaré distance calculation to capture hierarchical structure evolution; the implicit semantic association coding then mines the feature associations hidden in the difference features. In the double-layer modulation mechanism, the medical image difference semantic association coding matrix serves as a primary filter to eliminate interference features irrelevant to the current enhancement position in terms of semantics (such as abnormal responses corresponding to equipment artifacts); the medical image difference spatial modulation matrix serves as a secondary corrector to ensure that the enhancement process conforms to the spatial continuity constraint of the anatomical structure. This collaborative optimization enables the enhancement component to retain spatial irregularity while strengthening its functional connection semantic features with the surrounding areas.

[0033] Specifically, the sub-unit 1411 for extracting the differential channel features of the medical image to be enhanced is used to extract the channel feature vector at the (i,j) pixel position from the differential feature map of the medical image as the differential channel feature vector of the medical image to be enhanced, which is expressed by the formula for extracting the differential channel features of the medical image to be enhanced as:

[0034] F∈R H×W×C

[0035] v tbs =F(i,j,:)∈R C

[0036] where F is the differential feature map of the medical image, R is the set of real numbers, H and W are the height and width of each feature matrix of F along the channel dimension respectively, C is the number of channels of F, F(i,j,:) is the channel feature vector at the (i,j) pixel position in F, and v tbs is the differential channel feature vector of the medical image to be enhanced. It should be understood that considering the complexity of the lesion morphology in the medical image, such as the burr-like structure at the bifurcation of blood vessels or the irregular texture at the edge of the tumor, these features are likely to be blurred or misjudged under the fixed receptive field of the traditional rectangular convolution kernel. By selecting the channel feature vector at a specific pixel position (i,j), the system can focus on the subtle anatomical changes in the local area, such as the thickening of a certain bronchial wall or the signal abnormality of a certain area of white matter in the brain, thus avoiding the dilution of local pathological signals by global feature extraction. Among them, each channel feature vector essentially represents the response intensity of this pixel position in different semantic spaces (such as anatomical structure, tissue texture, pathological progression pattern). Extracting the differential channel feature vector of the medical image to be enhanced provides a basis for subsequent spatial-semantic joint enhancement. In this way, it is possible to achieve refined modeling of complex lesion morphologies through local feature selection, such as accurately capturing the microvascular infiltration characteristics of cirrhotic nodules in liver CT images.

[0037] Specifically, the sub-unit 1412 for randomly scanning the reference features of the medical image difference is used to perform n random scans on the differential feature map of the medical image to obtain n differential channel feature vectors of the medical image as a sparse set of reference features of the medical image difference, which is expressed by the formula for randomly scanning the reference features of the medical image difference as:

[0038] R c ={r1,r2,...,r i ,...,r n}

[0039] where R c is a sparse set of reference features of the medical image difference, r1, r2, r i and r nThe 1st, 2nd, ith, and nth medical image difference reference feature vectors in the sparse set of medical image difference reference feature vectors respectively, and r i ∈R C . It should be understood that by introducing a multiple random scanning strategy, the attentional mechanism of the human eye vision can be simulated, and the local representations of different anatomical structures can be dynamically captured in the medical image difference feature map, so as to form a "context" reference set. For example, in liver CT images, random sampling may cover different regions of liver parenchymal texture, vascular branch orientation, and lesion edges. This non-deterministic sampling essentially constructs a multi-scale and multi-orientation feature memory bank to balance the comprehensiveness of feature representation and computational efficiency. In this way, the medical image difference channel feature vectors obtained by each random scan not only contain spatial position information but also fuse multi-channel depth features (such as different activation patterns of convolutional layers), forming a sparse set of medical image difference reference feature vectors to avoid the overfitting risk caused by dense sampling. For example, in mammograms, the sensitivity to tumor edges will not be lost due to over-sampling of calcification points. Moreover, by retaining the key local patterns in the medical image difference feature map, cross-region context associations can be established, just like when analyzing brain MRI, the abnormal changes in the white matter fiber bundle orientation and cortical depression morphology can be associated simultaneously. In addition, by adopting sparse sampling instead of full coverage, the computational load can be significantly reduced while ensuring the information coverage, and its natural feature selection property helps to focus on the highly discriminative feature subset and avoid the interference of redundant or noisy features. This not only ensures the processing efficiency but also improves the quality of the analysis results.

[0040] Figure 5 is a block diagram of the explicit modeling modulation subunit in the deep learning-based intelligent auxiliary diagnosis and health management system for medical images according to an embodiment of the present application. As Figure 5As shown in the figure, in the embodiment of the present application, the explicit modeling modulation subunit 1413 is used to perform explicit modeling modulation on the sparse set of the to-be-enhanced medical image difference channel feature vectors and the medical image difference reference feature vectors to obtain the implicit coding vector of the medical image difference enhancement component, including: a medical image Poincaré distance calculation secondary subunit 1413-1, configured to calculate the Poincaré distance between each medical image difference reference feature vector in the sparse set of the to-be-enhanced medical image difference channel feature vectors and the medical image difference reference feature vectors to obtain a medical image difference spatial modulation matrix; a medical image implicit semantic association calculation secondary subunit 1413-2, configured to calculate the implicit semantic association between each medical image difference reference feature vector in the sparse set of the to-be-enhanced medical image difference channel feature vectors and the medical image difference reference feature vectors to obtain a set of medical image difference semantic association coding matrices; a difference feature enhancement component implicit coding secondary subunit 1413-3, configured to use each medical image difference semantic association coding matrix in the set of medical image difference semantic association coding matrices as a primary mask modulation unit and use the medical image difference spatial modulation matrix as a secondary mask modulation unit to perform explicit modeling modulation on the medical image difference information compensation coding vector between each medical image difference reference feature vector in the sparse set of the medical image difference reference feature vectors and the to-be-enhanced medical image difference channel feature vectors to obtain the implicit coding vector of the medical image difference enhancement component, where the medical image difference information compensation coding vector is the position-by-position difference vector between the medical image difference reference feature vector and the to-be-enhanced medical image difference channel feature vector.

[0041] Specifically, the medical image Poincaré distance calculation secondary subunit 1413-1 is configured to calculate the Poincaré distance between each medical image difference reference feature vector in the sparse set of the to-be-enhanced medical image difference channel feature vectors and the medical image difference reference feature vectors to obtain a medical image difference spatial modulation matrix, which is expressed by the medical image Poincaré distance calculation formula as:

[0042]

[0043] Where, ||·|| 2 is the square of calculating the Euclidean norm, arccosh is the inverse hyperbolic cosine function, M tbs-i is v tbs and r i the Poincaré distance between them, M tbs-1 …M tbs-n are the respective eigenvalues in the medical image difference spatial modulation matrix, M sIt is the modulation matrix of the medical image difference space. It should be understood that when dealing with complex lesions such as early tiny calcifications in pancreatic cancer and white matter plaques in multiple sclerosis, these lesions often exhibit multi-scale and multi-level heterogeneity characteristics. Due to its linear measurement characteristics, the traditional Euclidean distance is difficult to effectively model the tree-like hierarchical structure in the feature space. For example, in the MRI images of intraductal carcinoma of the breast, the characteristic distributions of cancerous tissues and surrounding adipose tissues may form a non-convex manifold structure. At this time, considering that the Poincaré distance is more suitable for capturing the data distribution of hierarchical or tree-like structures, the Poincaré distance can more accurately capture the progressive transition characteristics between the lesion edge and normal tissues through the measurement method of hyperbolic geometry. Calculating the Poincaré distance can construct a clinically interpretable spatial attention mechanism. By calculating the Poincaré distance between the feature vector of the difference channel of the medical image to be enhanced and each medical image difference reference feature vector in the sparse set of medical image difference reference feature vectors, the system essentially establishes a spatial semantic association network of difference features. Each element in the obtained modulation matrix of the medical image difference space not only quantifies the similarity between local features and the reference pattern but also implies the spatial relationships with pathological significance such as the growth direction of the lesion and the tissue infiltration path. For example, in the CT images of liver cancer, the modulation matrix of the medical image difference may show that the characteristics of the tumor core area and the metastatic foci around the portal vein are clustered at short distances, while maintaining a long distance from the normal liver parenchyma. This hierarchical distance distribution can directly guide the enhancement network to preferentially enhance the feature patterns related to tumor subtypes while suppressing the interference of vascular noise. In this way, the distance calculation based on hyperbolic geometry maps the originally high-dimensional and dense feature space into interpretable attention weights, enabling the system to focus on the feature enhancement of key pathological regions such as the spiculation signs at the edges of lung nodules and the edema zones around cerebral hemorrhage foci. Moreover, by capturing complementary features in the non-linear space, decoupled enhancement is achieved, solving the misjudgment problem caused by feature mixing in traditional methods.

[0044] Specifically, the secondary subunit 1413-2 for calculating the implicit semantic association of medical images is used to calculate the implicit semantic association between the feature vector of the difference channel of the medical image to be enhanced and each medical image difference reference feature vector in the sparse set of medical image difference reference feature vectors to obtain a set of medical image difference semantic association coding matrices, which is expressed by the medical image implicit semantic association calculation formula as:

[0045]

[0046] Where, is the matrix multiplication, W i is the weight matrix corresponding to r i T is the transpose operation, and d is W i and r iThe length of the vector after multiplication, and softmax is the softmax function. is v tbs and r i is the medical image difference semantic association coding matrix between them, that is, the i-th medical image difference semantic association coding matrix in the set of medical image difference semantic association coding matrices, M semantic is the set of medical image difference semantic association coding matrices. and are the 1st, 2nd, and nth medical image difference semantic association coding matrices in the set of medical image difference semantic association coding matrices respectively. It should be understood that when facing microscopic pathological changes such as early Alzheimer's disease brain amyloid deposition plaques and follicular structure variations in papillary thyroid carcinoma, traditional explicit association methods based on local texture or intensity features often fail. For example, amyloid plaques may present weak signals similar to perivascular spaces on MRI T2WI images. If only relying on pixel-level gray-scale similarity calculation, it is very easy to cause misjudgment. Through implicit semantic association, the co-variation pattern between "perivascular space edema" and "amyloid deposition" at the histopathological level can be mined through the non-linear mapping of the deep feature space. Even if there is an overlap in their imaging phenotypes, the essential differences can be revealed through the medical image difference semantic association coding matrix. Through implicit semantic association, a cognitive reasoning bridge with clinical interpretability can be constructed. Among them, each element of the medical image difference semantic association coding matrix not only quantifies the superficial similarity between the feature vector of the medical image difference channel to be enhanced and the sparse set of medical image difference reference feature vectors, but also decodes the deep pathological mechanism association through multi-layer non-linear transformation. For example, in the diagnosis of glioma WHO grade, the medical image difference semantic association coding matrix may show a strong semantic binding between the tumor heterogeneity region and the invasive growth pattern, while forming a significant separation from the simple edema area. This association modeling beyond the pixel level enables the system to understand that "spiculated edge" is not only an abstract description of the imaging feature, but also the projection of the pathological process of malignant lesion cell infiltration in the imaging space. In this way, through implicit semantic association, the traditional apparent matching paradigm based on low-order visual features can be broken through, and the metaphor mapping mechanism in cognitive linguistics can be used to extract abstract semantic representations from the deep structure of the data. This not only avoids redundant calculations at the original pixel level, but also reveals the deep cognitive laws obscured by the apparent features through the topological analysis of the latent semantic space, that is, the non-linear correspondence relationship between the essential attributes of the object and the observed features.

[0047] In an embodiment of the present application, the differential feature enhancement component implicit coding secondary subunit 1413-3 is configured to use each medical image differential semantic association coding matrix in the set of medical image differential semantic association coding matrices as a primary mask modulation unit and use the medical image differential spatial modulation matrix as a secondary mask modulation unit to perform explicit modeling modulation on the medical image differential information compensation coding vectors between each medical image differential reference feature vector in the sparse set of medical image differential reference feature vectors and the medical image differential channel feature vector to be enhanced to obtain a medical image differential enhancement component implicit coding vector. The medical image differential information compensation coding vector is the position-by-position difference vector between the medical image differential reference feature vector and the medical image differential channel feature vector to be enhanced, including: calculating the position-by-position difference vectors between the medical image differential channel feature vector to be enhanced and each medical image differential reference feature vector to obtain a set of medical image differential information compensation coding vectors; based on the primary mask modulation unit and the secondary mask modulation unit, performing coupled enhancement of the action of the gauge field covariant field on each medical image differential information compensation coding vector in the set of medical image differential information compensation coding vectors to obtain a set of medical image differential information compensation semantic space enhancement coding vectors; based on the primary mask modulation unit and the secondary mask modulation unit, performing explicit modeling modulation on the set of medical image differential information compensation semantic space enhancement coding vectors to obtain the medical image differential enhancement component implicit coding vector.

[0048] Specifically, the differential feature enhancement component implicit coding secondary subunit 1413-3 is represented by the differential feature enhancement component implicit formula as follows:

[0049] x i =r i -v tbs

[0050]

[0051] Wherein, x i is the medical image differential information compensation coding vector between v tbs and r i , x i ′ is the medical image differential information compensation semantic enhancement coding vector between v tbs and r i , x i ” is the medical image differential information compensation semantic space enhancement coding vector between v tbs and r i , n is the number of vectors in the sparse set of medical image differential reference feature vectors, and Δv tbsIt is the hidden coding vector of the medical image difference enhancement component. It should be understood that when there is noise interference or equipment parameter deviation in the medical image difference features to be enhanced, directly using the original medical image difference channel feature vector of the medical image to be enhanced may not accurately reflect the pathophysiological features of the lesion. Although the medical image difference reference feature vector contains prior knowledge, its global characteristics are prone to masking the local abnormal patterns of the lesion. Therefore, by constructing a two-level explicit modulation mechanism, the calculation of the medical image difference information compensation coding vector focuses on the spatial position difference between the medical image difference reference feature and the medical image difference channel feature of the medical image to be enhanced, rather than simply adding the eigenvalue. This differential vector quantization strategy can effectively isolate the pathological-related feature increment that needs to be compensated, such as the degree of tumor edge enhancement or the abnormal tissue density area, so as to achieve targeted enhancement. Through the gauge field-covariant field coupling mechanism, this secondary subunit regards the medical image difference semantic association coding matrix as the gauge field to constrain the global transformation invariance, ensuring that the enhancement process does not fail due to the overall deformation or rotation of the image; at the same time, it uses the medical image difference spatial modulation matrix as the covariant field to process the local curved space. Through the interaction coupling in the form of the partial derivative of the field space, the medical image difference information compensation coding dynamically adjusts the weight distribution of the semantic features while maintaining the continuity of the local anatomical structure. Specifically, it can first use to constrain the gauge transformation invariance, and then use to constrain the covariant transformation invariance. That is, by using partial differential operations in the field environment to represent the interaction strength between fields, the synergistic effect of the dynamic double-layer mask architecture and the adaptive information compensation mechanism is strengthened, thereby significantly improving the efficient coding ability of multi-dimensional information. This process realizes the explicit decoupling of spatial features through gradient field analysis, making the mask parameter optimization and the compensation amount adjustment form a closed-loop feedback, and then enhancing the fidelity and pattern discrimination of the signal representation. This double-layer mask modulation strategy first uses a semantic mask to screen out the anatomical regions associated with the target lesion, and then the spatial mask refines the compensation intensity to the pixel level accuracy, avoiding redundant calculations and ensuring the anatomical rationality of the enhancement effect. For example, in the enhancement of brain MRI, this mechanism can preferentially enhance the high-perfusion features of the tumor sub-region while suppressing the over-enhancement of normal white matter fibers, thereby improving the detection rate of small metastatic lesions.

[0052] Specifically, the medical image feature fusion subunit 1414 is used to fuse the hidden coding vector of the medical image difference enhancement component and the medical image difference channel feature vector of the medical image to be enhanced to obtain an enhanced medical image difference channel feature vector, where the enhanced medical image difference channel feature vector is the channel feature vector at the pixel position (i, j) of the enhanced medical image difference feature map, and is expressed by the enhanced medical image difference channel feature fusion formula as:

[0053] v enhanced = α·Δvtbs +β·v tbs

[0054] where α and β are weighted hyperparameters, and v enhanced is v tbs is the enhanced channel feature vector of the enhanced medical image difference, that is, the channel feature vector at the (i, j) pixel position of the enhanced medical image difference feature map. It should be understood that the enhanced channel feature vector of the medical image difference provides the morphological contour and local texture features of the lesion, ensuring the accurate anchoring of the anatomical structure, while the medical image difference enhancement component implicit coding vector injects the pathophysiological prior knowledge of the lesion, enabling the system to transcend the limitations of the apparent features and capture deep correlation patterns related to malignant tumors such as vascular distortion and edge enhancement. Fusing the two can endow it with semantic interpretability while retaining the localization ability of the original features, and also significantly improve the robustness of the features. When the image quality deteriorates, the system can rely on the pathological feature prior in the implicit coding for compensatory analysis, thereby reducing the dependence on high-quality image acquisition devices while maintaining the diagnostic accuracy rate, and finally achieving effective enhancement at the feature level and completing the entire feature-level enhancement process based on spatial-semantic joint explicit modeling.

[0055] In an embodiment of the present application, the generating unit 142 is configured to: input the enhanced medical image difference feature map into a heat map generator based on AIGC to obtain the visualized difference heat map. It should be understood that the human visual system is more sensitive to changes in color and brightness. Compared with abstract feature data, visualized images are easier to understand and interpret. Presenting medical image differences in the form of a heat map can utilize the advantages of human vision, enabling doctors to more intuitively perceive the differences between images and conforming to doctors' visual cognitive habits. That is to say, the visualized difference heat map is a method that converts difference features (such as lesion changes, tissue density differences, etc.) in medical image data into an intuitive and visible image through color coding and intensity mapping techniques. Its core purpose is to convert complex numerical difference data into visual signals that can be quickly recognized by humans, and the color gradient is more in line with the visual perception characteristics of the human eye. The visualized difference heat map can clearly display the differences between the current medical image and the historical medical image. Doctors can quickly locate the areas with larger differences by observing the color and brightness distribution of the heat map, and thus conduct a more targeted detailed analysis of these areas, assisting doctors in making more accurate diagnostic decisions and improving diagnostic efficiency. In a specific embodiment of the present application, in the heat map generator based on AIGC technology, the input enhanced medical image difference feature map is first standardized to ensure that all values fall within a preset range. Subsequently, using a pre-defined color mapping scheme, the heat map generator based on AIGC technology begins to convert the value corresponding to each pixel point in the enhanced medical image difference feature map into a specific color intensity. Specifically, by converting the differences in the patient's medical images at different time points (such as lesion volume, density, and morphological changes) into a color-coded visual overlay layer, the spatial distribution and intensity of the changed areas are intuitively displayed to assist doctors in quickly locating key areas. For example, in a common red-blue color scheme, red is usually used to represent higher values or significant changes, while blue represents lower values or relatively stable states. Through this precise mapping between color and value, even subtle pathological changes can be clearly visible on the finally generated heat map. In addition, to further enhance the interpretability value of the heat map, the heat map generator based on AIGC technology also adjusts the smoothness and contrast of the color gradient according to the characteristics of the human visual system. This not only helps to improve the highlighting effect of the lesion area but also makes it easier for doctors to identify small lesions that may be overlooked. During the entire conversion process, the application of AIGC technology ensures that the conversion from the enhanced medical image difference feature map to the heat map not only retains the core information of the original data but also adds visual elements that are easy to understand and analyze. Finally, after a series of fine-tuning and optimization, the finally generated visualized difference heat map will be presented to the user in an intuitive and expressive form.This heat map can not only accurately reflect the differences between the current medical image and historical medical images, but also reveal the development trend of lesions and their anatomical distribution through the variation of color shades, thus greatly assisting radiologists in making more accurate diagnostic decisions. This process fully demonstrates the great potential of the combination of modern information technology and medical image analysis.

[0056] In summary, the intelligent assisted diagnosis and health management system 100 for medical images based on deep learning according to the embodiments of the present application is elucidated. First, it preprocesses the image data to eliminate device noise and physiological artifact interference, and extracts multi-scale spatial-semantic features of cross-temporal images. It captures the lesion evolution trajectory using feature map difference operations, then strengthens pathological features and suppresses redundancy of the differential features through spatial-semantic joint enhancement. Finally, it converts the temporal differential features into a color gradient map that conforms to the cognitive habits of radiologists with the help of a generation mechanism. In this way, it can overcome misjudgments caused by motion artifacts and physiological changes, enhance the distinguishability of minor pathological changes while maintaining the spatial topological structure of the lesion area, so as to both enhance the expression of minor but crucial pathological features and intuitively map them to the anatomical sites of the original image.

[0057] As described above, the intelligent assisted diagnosis and health management system 100 for medical images based on deep learning according to the embodiments of the present application can be implemented in various terminal devices. In one example, the intelligent assisted diagnosis and health management system 100 for medical images based on deep learning can be integrated into the terminal device as a software module and / or a hardware module. For example, the intelligent assisted diagnosis and health management system 100 for medical images based on deep learning can be a software module in the operating system of the terminal device, or can be an application program developed for the terminal device; of course, the intelligent assisted diagnosis and health management system 100 for medical images based on deep learning can also be one of the many hardware modules of the terminal device.

[0058] Alternatively, in another example, the intelligent assisted diagnosis and health management system 100 for medical images based on deep learning and the terminal device can also be separate devices, and the intelligent assisted diagnosis and health management system 100 for medical images based on deep learning can be connected to the terminal device through wired and / or wireless networks and transmit and interact information in accordance with a predefined data format.

[0059] In several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division, and there can be other division methods in actual implementation.

[0060] The module described as a separation component may or may not be physically separated. The component shown as a module may or may not be a physical unit, that is, it may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0061] In addition, in each embodiment of the present invention, each functional module can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated unit can be implemented in the form of hardware or in the form of a combination of hardware and software functional modules.

[0062] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms.

[0063] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be construed as limiting the claimed invention.

[0064] In addition, it is obvious that the word "including" does not exclude other units or steps, and the singular does not exclude the plural. The multiple units or devices stated in the system claims can also be implemented by one unit or device through software or hardware. Words such as "second" are used to denote names and do not denote any particular order.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit of the technical solutions of the present invention.

Claims

1. An intelligent auxiliary diagnosis and health management system for medical images based on deep learning, characterized in that, Including: A medical image data acquisition module, configured to acquire current medical image data and historical medical image data of a patient object; A preprocessing module, configured to perform image preprocessing on the current medical image data and the historical medical image data to obtain preprocessed current medical image data and preprocessed historical medical image data; A difference feature calculation module, configured to perform difference calculation based on image features on the preprocessed current medical image data and the preprocessed historical medical image data to obtain a medical image difference feature map; A difference enhancement mapping module, configured to perform difference feature enhancement and color mapping on the medical image difference feature map to obtain a visualized difference heat map, wherein the difference enhancement mapping module includes: an enhancement unit, configured to perform image enhancement on the medical image difference feature map based on feature space-semantic two-dimensional constraints to obtain an enhanced medical image difference feature map; a generation unit, configured to perform mapping on the enhanced medical image difference feature map to obtain the visualized difference heat map.

2. The intelligent assisted diagnosis and health management system for medical images based on deep learning according to claim 1, wherein The historical medical image data is the data collected in the previous acquisition of the current medical image data.

3. The intelligent auxiliary diagnosis and health management system for medical images based on deep learning according to claim 2, characterized in that, The image preprocessing includes denoising filtering, artifact correction, contrast enhancement, and normalization processing.

4. The intelligent auxiliary diagnosis health management system for medical images based on deep learning according to claim 3, characterized in that, The difference feature calculation module is configured to: Use an image feature extractor based on serpentine convolution to respectively extract features from the preprocessed current medical image data and the preprocessed historical medical image data to obtain a current medical image data feature map and a historical medical image data feature map; Calculate the position-wise difference between the current medical image data feature map and the historical medical image data feature map to obtain the medical image difference feature map.

5. The intelligent auxiliary diagnosis and health management system for medical images based on deep learning according to claim 4, characterized in that The enhancement unit includes: A medical image difference channel feature extraction sub-unit for extracting the channel feature vector at the (i,j) pixel position from the medical image difference feature map as the medical image difference channel feature vector to be enhanced; A medical image difference reference feature random scanning sub-unit for performing n random scans on the medical image difference feature map to obtain n medical image difference channel feature vectors as a sparse set of medical image difference reference feature vectors; A dominant modeling modulation sub-unit for performing dominant modeling modulation on the medical image difference channel feature vector to be enhanced and the sparse set of medical image difference reference feature vectors to obtain a medical image difference enhancement component implicit coding vector; A medical image feature fusion sub-unit for fusing the medical image difference enhancement component implicit coding vector and the medical image difference channel feature vector to be enhanced to obtain an enhanced medical image difference channel feature vector, wherein the enhanced medical image difference channel feature vector is the channel feature vector at the pixel position (i,j) of the enhanced medical image difference feature map.

6. The intelligent auxiliary diagnosis and health management system for medical images based on deep learning according to claim 5, characterized in that, The dominant modeling modulation sub-unit includes: The Poincaré distance calculation secondary subunit for medical images is used to calculate the Poincaré distance between the feature vector of the difference channel of the medical image to be enhanced and each reference feature vector of the sparse set of reference feature vectors of the medical image differences to obtain the modulation matrix of the medical image difference space; The implicit semantic association calculation secondary subunit for medical images is used to calculate the implicit semantic association between the feature vector of the difference channel of the medical image to be enhanced and each reference feature vector of the sparse set of reference feature vectors of the medical image differences to obtain a set of encoding matrices of the medical image difference semantic association; The implicit encoding secondary subunit of the differential feature enhancement component is used to use each encoding matrix of the medical image difference semantic association in the set of encoding matrices of the medical image difference semantic association as the primary mask modulation unit and the modulation matrix of the medical image difference space as the secondary mask modulation unit to perform explicit modeling modulation on the medical image difference information compensation encoding vector between each reference feature vector of the sparse set of reference feature vectors of the medical image differences and the feature vector of the difference channel of the medical image to be enhanced to obtain the implicit encoding vector of the medical image difference enhancement component, and the medical image difference information compensation encoding vector is the position-by-position difference vector between the reference feature vector of the medical image difference and the feature vector of the difference channel of the medical image to be enhanced.

7. The intelligent assisted diagnosis and health management system for medical images based on deep learning according to claim 6, characterized in that The implicit encoding secondary subunit of the differential feature enhancement component is used for: Calculating the position-by-position difference vector between the feature vector of the difference channel of the medical image to be enhanced and each reference feature vector of the medical image differences to obtain a set of medical image difference information compensation encoding vectors; Based on the primary mask modulation unit and the secondary mask modulation unit, performing coupled enhancement of the action of the gauge field covariant field on each medical image difference information compensation encoding vector in the set of medical image difference information compensation encoding vectors to obtain a set of enhanced encoding vectors of the medical image difference information compensation semantic space; Based on the primary mask modulation unit and the secondary mask modulation unit, performing explicit modeling modulation on the set of enhanced encoding vectors of the medical image difference information compensation semantic space to obtain the implicit encoding vector of the medical image difference enhancement component.

8. The intelligent auxiliary diagnosis and health management system for medical images based on deep learning according to claim 7, characterized in that, The generating unit is used for: inputting the enhanced medical image difference feature map into the heat map generator based on AIGC to obtain the visualized difference heat map.

Images