Computer-aided diagnosis method based on deep learning

Abstract

The invention discloses a computer-aided diagnosis method based on deep learning, and the method comprises the steps: collecting medical image data, and carrying out the preprocessing of the medical image data, and obtaining standardized medical image data; constructing an improved MoCo model to obtain a high-dimensional diagnosis feature vector; embedded space clustering consistency optimization is executed, and a diagnosis feature vector subjected to space structure mapping processing is obtained; calculating a correlation measure between each feature dimension and the diagnosis tag, and calculating a redundancy measure between each feature dimension; executing a Boruta algorithm to obtain an optimal diagnosis feature subset; and extracting corresponding feature dimensions to form to-be-diagnosed feature sub-vectors, and generating a diagnosis result. According to the method, the improved MoCo model and the Boruta algorithm are introduced, so that deep extraction, accurate screening and high-reliability computer-aided diagnosis of medical image diagnosis features are realized.

Description

Technical Field

[0001] This invention relates to the field of computer-aided medical diagnosis technology, and in particular to a computer-aided diagnosis method based on deep learning. Background Technology

[0002] Currently, with the continuous improvement of medical intelligence and informatization, intelligent technology plays an increasingly important role in clinical diagnosis, and computer-aided diagnosis is becoming a key direction for the development of smart healthcare. In traditional technologies, the analysis and diagnosis of medical images typically rely on single judgment methods, such as manual image reading experience, image analysis based on static rules, or traditional statistical discrimination methods. These methods struggle to accurately address the complex variations caused by different equipment parameters, lesion types, and imaging noise. Because lesion manifestations are highly uncertain and variable, diagnostic processes relying on fixed thresholds or traditional pattern recognition methods are easily affected by subjective factors, have limited defect identification capabilities, and are prone to misdiagnosis and missed diagnosis, failing to meet the clinical demand for high-precision, widely applicable disease screening.

[0003] Existing technologies are beginning to combine deep feature extraction with intelligent diagnostic modeling, but they mostly focus on learning single structural features or representing fixed-dimensional features, making it difficult to take into account both the local texture information and the overall structural relationship of small lesions, resulting in incomplete feature representation. While some studies employing self-supervised learning or feature selection techniques have improved label dependence and feature utilization, they still suffer from technical problems such as rigid coding structures, insufficient feature distribution representation capabilities, and unstable feature selection. Existing technologies struggle to adaptively handle complex semantic differences in images, and existing feature selection methods are not robust enough in identifying redundant features in high-dimensional medical imaging scenarios, making it difficult to fully realize the diagnostic value of high-dimensional feature sets and limiting the widespread application of intelligent medical image diagnostic technology in real-world clinical settings.

[0004] Therefore, how to provide a computer-aided diagnostic method based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based computer-aided diagnostic method. This invention utilizes an intelligent diagnostic strategy combining unsupervised contrastive learning and feature selection techniques, fully leveraging an improved MoCo model and the Boruta algorithm to automatically extract and select diagnostic features from medical image data. The invention details the application of the improved MoCo model with a heterogeneous coding structure to medical image feature representation. High-dimensional diagnostic representations are obtained through joint encoding of local and global features. Then, by performing embedding space clustering mapping on the high-dimensional features, unlabeled features are transformed into representations with separable diagnostic significance. This invention introduces the Boruta algorithm based on structure-preserving perturbation and dynamic threshold removal to achieve stable extraction of key diagnostic dimensions. Based on the selected features, a distribution comparison and category matching mechanism is constructed to output the final diagnostic results. This invention possesses advantages such as requiring minimal manual annotation, high diagnostic accuracy, strong model stability, and adaptability to various medical image types, contributing to improved reliability and scalability of intelligent clinical auxiliary diagnosis.

[0006] A computer-aided diagnostic method based on deep learning according to an embodiment of the present invention includes: Collect medical imaging data, preprocess the medical imaging data, and obtain standardized medical imaging data; Based on standardized medical image data, an improved MoCo model is constructed. The first encoder and the second encoder of the improved MoCo model adopt a heterogeneous network structure to obtain a high-dimensional diagnostic feature vector. Based on the high-dimensional diagnostic feature vector, embedding space clustering consistency optimization is performed to minimize the embedding distance of samples of the same class and maximize the embedding distance of samples of different classes. Spatial structure mapping is then performed on the high-dimensional diagnostic feature vector to obtain the diagnostic feature vector processed by spatial structure mapping. Based on the diagnostic feature vector processed by spatial structure mapping, the correlation measure between each feature dimension and the diagnostic label and the redundancy measure between each feature dimension are calculated. Based on correlation and redundancy measures, the Boruta algorithm is used to construct a shadow feature set with the same number of original feature dimensions. The importance of the original feature dimensions and the shadow feature dimensions is statistically tested, and unimportant feature dimensions are iteratively eliminated to obtain the optimal diagnostic feature subset. Based on the optimal diagnostic feature subset, the corresponding feature dimensions are extracted from the diagnostic feature vectors processed by spatial structure mapping to form the feature subvectors to be diagnosed, and feature distribution comparison analysis is performed to generate diagnostic results.

[0007] Optionally, the medical imaging data specifically includes X-ray imaging data, CT imaging data, MRI imaging data, nuclear medicine imaging data, and ultrasound imaging data.

[0008] Optionally, the preprocessing of medical image data specifically includes grayscale normalization, noise suppression and structure-preserving filtering, artifact correction, automatic extraction of regions of interest, spatial geometric correction, and resolution resampling.

[0009] Optionally, obtaining the high-dimensional diagnostic feature vector includes: An improved MoCo model is constructed, which consists of a heterogeneous query encoder module, a heterogeneous momentum encoder module, a dynamic momentum update module, a semantic alignment feature queue module, and a contrastive loss calculation module. Based on standardized medical image data, a first data augmentation operation and a second data augmentation operation are performed respectively to form a first view and a second view; The first view is input into the heterogeneous query encoder module, and variable convolution kernel processing and dual-path downsampling processing are performed on the first view to form downsampling features at different scales. Local stability enhancement feature fusion processing is then performed, and a differentiated fusion strategy is adopted for smooth and variable regions to obtain query feature representations. The second view is input into the heterogeneous momentum encoder module for adaptive patch division. The patch size is adjusted according to the structural complexity of the image region. The patch sequence is converted into a semantic representation. An attention guidance mechanism is introduced to adaptively adjust the attention weight allocation according to the gray density difference of different regions. Then, a feedforward transformation process for global consistency reconstruction is performed to obtain the key feature representation by reconstructing the structural relationship between patches. The parameters of the heterogeneous query encoder module are obtained by the dynamic momentum update module as the target parameters. A momentum adjustment strategy is introduced to adaptively adjust the momentum update coefficient based on the difference between the parameters of the heterogeneous momentum encoder module and the target parameters. The current update amplitude is determined based on the momentum update coefficient, and the parameters of the heterogeneous momentum encoder module are adjusted to gradually approach the target parameters. Input the key feature representation into the semantic alignment feature queue module, delete the earliest key feature representation that entered the queue, and perform replacement processing based on the semantic difference between the key feature representation and the key feature representation in the queue to obtain the updated semantic alignment feature queue. The contrastive loss calculation module performs contrastive learning operations based on the query feature representation, key feature representation, and updated semantic alignment feature queue, and introduces a semantic interval control mechanism to dynamically adjust the contrastive learning scale to generate a high-dimensional diagnostic feature vector.

[0010] Optionally, obtaining the diagnostic feature vector processed by spatial structure mapping includes: Embedding mapping is performed on the high-dimensional diagnostic feature vectors of each medical image sample to convert the high-dimensional diagnostic feature vectors into corresponding embedding vectors, forming the initial embedding vectors. Based on the initial embedding vector and the diagnostic labels corresponding to each medical image sample, the initial embedding vectors belonging to the same diagnostic category are aggregated to obtain the category representation of each diagnostic category, and a sample constraint relationship is constructed in which samples of the same category are close to each other and samples of different categories are separated from each other. Based on the sample constraint relationship, clustering consistency optimization is performed on the initial embedding vector in the embedding space. The position of each embedding vector in the embedding space is adjusted, the distance between embedding vectors belonging to the same diagnostic category is gradually reduced, and the distance between embedding vectors belonging to different diagnostic categories is gradually increased, so as to obtain the optimized embedding vector. The optimized embedding vectors are subjected to spatial structure mapping processing. Based on the distribution structure of the optimized embedding vectors in the embedding space, a new diagnostic feature representation is formed, resulting in a diagnostic feature vector processed by spatial structure mapping.

[0011] Optionally, the calculation of the correlation metric between each feature dimension and the diagnostic label, and the redundancy metric between each feature dimension, includes: The values ​​of each medical image sample in each feature dimension of the diagnostic feature vector processed by spatial structure mapping are summarized to form the corresponding feature dimension sequence, and the known diagnostic category information of each medical image sample is obtained to form a diagnostic category sequence. For each feature dimension, a correlation measure between the feature dimension and the diagnostic category is calculated based on the joint distribution relationship between the feature dimension sequence and the diagnostic category sequence. For any two different feature dimensions, calculate the redundancy measure between the two feature dimensions based on the changing trend relationship between the corresponding feature dimension sequences.

[0012] Optionally, obtaining the optimal diagnostic feature subset includes: Based on correlation and redundancy measures, and combined with the diagnostic feature vector processed by spatial structure mapping, the feature dimensions in the diagnostic feature vector are collected to construct a set of original feature dimensions to be screened, and a corresponding importance record entry is established for each original feature dimension. The Boruta algorithm is executed. For each original feature dimension in the set of original feature dimensions to be screened, a structure-preserving perturbation mechanism is introduced to perturb the value distribution of the original feature dimension while keeping the local structural relationship unchanged, generating a corresponding shadow feature dimension. The number of shadow feature dimensions is consistent with the number of original feature dimensions, forming a shadow feature set. For each original feature dimension, the category discrimination score is calculated based on the differences in the value distribution in different diagnostic categories, and the corresponding shadow category discrimination score is calculated for each shadow feature dimension in the shadow feature set in the same way; For each original feature dimension, the intra-class consistency change is calculated based on the value fluctuation within each diagnostic category, and the corresponding intra-class consistency change is calculated for each shadow feature dimension in the shadow feature set in the same way. The importance statistics of the original features are constructed based on the category discrimination score and intra-class consistency change of the original feature dimension. The importance benchmark of the shadow feature is constructed based on the corresponding statistics of the shadow feature dimension. A dynamic threshold elimination mechanism is introduced to retain or eliminate features based on whether the importance statistics of the original features are significantly higher than the importance benchmark of the shadow features. In each iteration, the corresponding shadow feature dimension is regenerated for the retained original feature dimension, and the calculation of class discrimination score, intra-class consistency change amount and dynamic elimination are repeatedly performed. An adaptive termination strategy is introduced to determine whether to terminate the iteration based on whether the change in the number of retained original features has reached a stable state. After the iteration terminates, the set of original feature dimensions that are retained is taken as the optimal diagnostic feature subset.

[0013] Optionally, generating diagnostic results includes: Based on the optimal diagnostic feature subset, the corresponding feature dimension index set is determined, and a correspondence is established between the feature dimension index set and the diagnostic feature vector processed by spatial structure mapping. Based on the feature dimension index set, the values ​​of the corresponding feature dimensions are extracted from the diagnostic feature vector corresponding to each medical image sample after spatial structure mapping, forming the corresponding feature sub-vector to be diagnosed. Based on the diagnostic feature subvectors of medical image samples and the corresponding known diagnostic category information, the feature distribution of the diagnostic feature subvectors is statistically analyzed for each diagnostic category to form a reference feature distribution set for each diagnostic category. The diagnostic feature subvectors of the medical image samples to be diagnosed are compared with the reference feature distribution set. The distribution difference measure between the vectors and the reference feature distributions of each diagnostic category is calculated, and the corresponding category matching degree set is generated. The target diagnostic category corresponding to the medical image sample to be diagnosed is determined based on the category matching degree set, and the target diagnostic category is output as the diagnostic result.

[0014] The beneficial effects of this invention are: This invention proposes a deep learning-based computer-aided diagnostic method that employs a diagnostic strategy that integrates unsupervised contrastive learning and feature selection. This method achieves feature extraction and diagnostic recognition of medical images without requiring extensive manual annotation. By constructing an improved MoCo model, this invention adopts a heterogeneous structure design for the query encoder and momentum encoder, introducing convolutional extraction and self-attention feature representation methods adapted to different image regions. Furthermore, through dynamic momentum updates and semantically aligned feature queues, it enables high-dimensional image features to possess stable representation capabilities in both local texture and global structure, effectively overcoming the problems of traditional isomorphic networks' feature bias and incompatibility with complex anatomical differences. This invention further introduces embedded spatial clustering mapping on top of the high-dimensional representation, endowing unlabeled features with separable diagnostic attributes and improving the discriminative power of image representation.

[0015] In the feature selection stage, this invention utilizes an improved Boruta algorithm, combined with a structure-preserving perturbation strategy, differentiated category discrimination scoring, intra-class consistency analysis, and dynamic threshold elimination, to achieve importance assessment without relying on the model. This ensures that the obtained feature subset can stably represent the differential features between different disease categories, effectively improving the reliability and compactness of the selection process. Finally, diagnostic results are output based on the comparison of the selected feature distribution and the category matching mechanism, making the diagnostic process more interpretable and stable.

[0016] The method of this invention realizes a diagnostic path that combines unsupervised learning with robust feature selection. It has the advantages of not requiring a lot of manual annotation, high feature recognition accuracy, strong adaptability to multiple types of medical images, and good interpretability of diagnostic results. It can provide more reliable data support and application value for clinical image-assisted diagnosis. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0018] Figure 1 This is a flowchart of a computer-aided diagnostic method based on deep learning proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved MoCo model of a deep learning-based computer-aided diagnostic method proposed in this invention. Figure 3 This is a flowchart illustrating the Boruta algorithm execution of a deep learning-based computer-aided diagnostic method proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 , Figure 2 and Figure 3 A computer-aided diagnostic method based on deep learning, comprising: Collect medical imaging data, preprocess the medical imaging data, and obtain standardized medical imaging data; Based on standardized medical image data, an improved MoCo model is constructed. The first encoder and the second encoder of the improved MoCo model adopt a heterogeneous network structure to obtain a high-dimensional diagnostic feature vector. Based on the high-dimensional diagnostic feature vector, embedding space clustering consistency optimization is performed to minimize the embedding distance of samples of the same class and maximize the embedding distance of samples of different classes. Spatial structure mapping is then performed on the high-dimensional diagnostic feature vector to obtain the diagnostic feature vector processed by spatial structure mapping. Based on the diagnostic feature vector processed by spatial structure mapping, the correlation measure between each feature dimension and the diagnostic label and the redundancy measure between each feature dimension are calculated. Based on correlation and redundancy measures, the Boruta algorithm is used to construct a shadow feature set with the same number of original feature dimensions. The importance of the original feature dimensions and the shadow feature dimensions is statistically tested, and unimportant feature dimensions are iteratively eliminated to obtain the optimal diagnostic feature subset. Based on the optimal diagnostic feature subset, the corresponding feature dimensions are extracted from the diagnostic feature vectors processed by spatial structure mapping to form the feature subvectors to be diagnosed, and feature distribution comparison analysis is performed to generate diagnostic results.

[0021] In this embodiment, the medical imaging data specifically includes X-ray imaging data, CT imaging data, MRI imaging data, nuclear medicine imaging data, and ultrasound imaging data.

[0022] In this embodiment, the preprocessing of medical image data specifically includes grayscale normalization, noise suppression and structure-preserving filtering, artifact correction, automatic extraction of regions of interest, spatial geometric correction, and resolution resampling.

[0023] In this embodiment, obtaining the high-dimensional diagnostic feature vector includes: An improved MoCo model is constructed, comprising a heterogeneous query encoder module, a heterogeneous momentum encoder module, a dynamic momentum update module, a semantic alignment feature queue module, and a contrastive loss calculation module. Specifically, the construction of the improved MoCo model involves: The query encoder of the MoCo model is replaced with a heterogeneous query encoder module that uses a convolutional neural network structure and can perform multi-scale local feature extraction. The momentum encoder is replaced with a heterogeneous momentum encoder module that uses a hierarchical attention structure and can represent global structural information. The query encoder and momentum encoder in the MoCo model are improved to encoders with different structures. The original fixed momentum update method of the MoCo model is extended by adding a dynamic momentum update module, which is connected after the heterogeneous momentum encoder module. Dynamic adjustment is performed according to the parameter differences between the heterogeneous query encoder module and the heterogeneous momentum encoder module to perform adaptive update of the heterogeneous momentum encoder module. A spike semantic difference is added to the feature queue structure of the MoCo model to update the queue, forming a semantic alignment feature queue module. A semantic interval adjustment mechanism is added to the contrastive loss structure of the MoCo model to form a contrastive loss calculation module. Based on standardized medical image data, a first data augmentation operation and a second data augmentation operation are performed respectively to form a first view and a second view. Specifically, the first and second data augmentation operations are performed as follows: The first data augmentation operation generates a first view by performing rotation, translation, scaling, contrast adjustment, and brightness perturbation on the medical image; the second data augmentation operation generates a second view by performing flipping, cropping, noise injection, blurring, and grayscale perturbation on the medical image. The first view is input into the heterogeneous query encoder module, where variable convolution kernel processing and dual-path downsampling are performed to generate downsampled features at different scales. Local stability-enhancing feature fusion processing is then performed, employing a differentiated fusion strategy for smooth and variable regions to obtain the query feature representation, where: The process of performing variable convolution kernel processing and dual-path downsampling on the first view is as follows: In the feature extraction stage, convolution operations are performed sequentially on the first view. The size, shape and sampling method of different convolution kernels are different to obtain local response features at multiple scales. Based on the features processed by the variable convolution kernel, two independent downsampling branches are constructed. One branch uses a fixed stride sampling method to downsample the local response features, and the other branch uses a region-sensitive non-uniform sampling method to downsample the local response features, forming downsampled features at different scales. Differentiated fusion strategy refers to using different feature combination methods for different types of regions based on the differences in structural distribution between smooth and variable regions in the downsampled features. When the features obtained from the two downsampled paths are fused, different fusion methods are used according to the region category to form a fusion feature structure that can distinguish different region types. The feature fusion process that enhances local stability is specifically as follows: The downsampled features obtained by dual-path downsampling are subjected to local continuity detection in each region to distinguish between smooth regions and changing regions. Different feature combination methods are used for smooth regions and changing regions respectively. Feature aggregation based on continuity weight is performed on smooth regions, and feature superposition based on structural differences is performed on changing regions. The downsampled features output by the two downsampling paths are weighted and synthesized to generate a query feature representation that includes regional difference fusion results. The second view is input into the heterogeneous momentum encoder module for adaptive patch partitioning. The patch size is adjusted according to the structural complexity of the image region. The patch sequence is converted into a semantic representation. An attention-guided mechanism is introduced to adaptively adjust the attention weight allocation based on the gray-level density differences in different regions. A feedforward transformation process for global consistency reconstruction is then performed. Key feature representations are obtained by reconstructing the structural relationships between patches, where: The adaptive patch division is specifically performed as follows: Structural complexity analysis is performed on each local region in the second view. Based on the changes in grayscale variation, edge density, and texture richness of each local region, the local regions are divided into different complexity levels. The greater the rate of change, the higher the complexity level. Regions with high complexity levels are divided by selecting a patch size smaller than the patch size baseline, while regions with low complexity levels are divided by selecting a patch size larger than the patch size baseline. The corresponding local regions are then segmented according to the patch size to generate a patch sequence composed of patches of different sizes. Attention guidance mechanism refers to dividing patches into multiple weight levels based on the gray density statistics of each patch in the patch sequence, and assigning different attention weights to each patch. The larger the gray density statistics, the greater the attention weight, thus forming a patch sequence with differentiated weight allocation. The feedforward transformation process for performing global consistency reconstruction is specifically as follows: The patch sequence with attention weights is sequentially input into multiple feedforward transformation units in hierarchical order. The patch sequence is subjected to dimensional and numerical transformations and recombined to form a new feature arrangement structure. The correlation information between patches is combined according to their positional relationship, and the adjacency, order, and spatial arrangement relationships between patches are reconstructed. The sequential connection method between patches is determined according to their adjacent positions, and the spatial connection method is determined according to their arrangement layout. Adjacent features between patches are reconnected to form a structural relationship corresponding to the patch layout. The structural connection information between patches in the patch sequence is updated after each layer of feedforward transformation. Finally, the patch feature representation with consistent structural relationships between patches is output as the key feature representation. The parameters of the heterogeneous query encoder module are obtained by the dynamic momentum update module as the target parameters. A momentum adjustment strategy is introduced to adaptively adjust the momentum update coefficients based on the difference between the parameters of the heterogeneous momentum encoder module and the target parameters. The current update amplitude is determined based on the momentum update coefficients, and the parameters of the heterogeneous momentum encoder module are adjusted to gradually approach the target parameters. The parameters of the heterogeneous query encoder module refer to the network structure parameters that perform convolution operations, downsampling processing, and local stability enhancement feature fusion processing, including convolution kernel weights, channel weights, and fusion weights. The parameters of the heterogeneous momentum encoder module refer to the network structure parameters that perform adaptive patch partitioning, attention weight allocation, and global consistency reconstruction feedforward transform processing, including patch feature encoding weights, attention weight parameters, and feedforward transform unit weights. The momentum adjustment strategy refers to assigning a corresponding momentum update coefficient to the parameter update based on the degree of difference between the parameters after the heterogeneous query encoder module and the heterogeneous momentum encoder module generate their respective parameters. The greater the parameter difference, the higher the assigned momentum update coefficient, and the smaller the parameter difference, the lower the assigned momentum update coefficient. The momentum update coefficient used in each iteration corresponds to the parameter difference. The determination of the current update magnitude based on the momentum update coefficient is specifically as follows: The difference between the current parameters and the target parameters of the heterogeneous momentum encoder module is calculated, and the difference is scaled according to the assigned momentum update coefficient to obtain the corresponding update amplitude. The adjustment of the parameters of the heterogeneous momentum encoder module is specifically as follows: The determined current update magnitude is superimposed with the current parameters of the momentum encoder module. The superimposed parameter value is updated in the direction of the target parameter, and the superposition update operation is repeated in each iteration. The key feature representation is input into the semantic alignment feature queue module. The earliest key feature representation in the queue is deleted, and a replacement process is performed based on the semantic difference between the key feature representation and the key feature representations in the queue, resulting in an updated semantic alignment feature queue. The replacement process based on the semantic difference between the key feature representation and the key feature representations in the queue specifically involves: Semantic distance calculation is performed on the current key feature representation and each existing key feature representation in the semantic alignment feature queue. The difference values ​​of the corresponding feature dimensions of the current key feature representation and the key feature representation in the semantic alignment feature queue are extracted. The differences of each dimension are accumulated and then normalized and compressed to form a distance value. The semantic difference degree is determined according to the size of the distance value. Key feature representations in the semantic alignment feature queue with a semantic difference degree less than the difference threshold are marked as replaceable objects and are replaced in order of semantic difference degree from small to large. After the update within the replacement mark range is completed, the current key feature representation is inserted into the replacement position to form the updated semantic alignment feature queue. The contrastive loss calculation module performs contrastive learning operations based on the query feature representation, key feature representation, and the updated semantic alignment feature queue. It also introduces a semantic margin adjustment mechanism to dynamically adjust the contrastive learning scale, generating a high-dimensional diagnostic feature vector, where: The execution of the comparison learning operation is specifically as follows: A similarity measure is performed on the query feature representation and the key feature representation, and the similarity is used as a positive sample measure. A difference measure is performed on the query feature representation and the features in the updated semantically aligned feature queue, and the difference is used as a negative sample measure. The positive sample measure and the negative sample measure are weighted and combined to form a contrastive learning statistic. The semantic interval control mechanism refers to assigning different interval coefficients based on the similarity between the query feature representation and the key feature representation during the calculation of the contrastive learning statistic. When the similarity is higher than the similarity threshold, an interval coefficient lower than the baseline is assigned; when the similarity is lower than the similarity threshold, an interval coefficient higher than the baseline is assigned. The interval coefficient serves as a control factor for the contrastive learning statistic, dynamically adjusting the contrast scale between samples in a weighted manner. The generation of the high-dimensional diagnostic feature vector specifically involves: Numerical mapping is performed on the contrastive learning statistics adjusted by the semantic interval control mechanism in each dimension. Each contrastive learning statistic value is converted into a corresponding embedding expression. The embedding expression is then expanded and stretched in dimension. Encoding transformation is performed on the expanded embedding expression. Each dimension of the embedding expression is numerically transformed and positionally encoded according to the encoding parameters. Finally, a multidimensional vector after numerical mapping, dimensional expansion and encoding transformation is output, which serves as a high-dimensional diagnostic feature vector.

[0024] In this embodiment, obtaining the diagnostic feature vector processed by spatial structure mapping includes: An embedding mapping process is performed on the high-dimensional diagnostic feature vectors of each medical image sample to convert the high-dimensional diagnostic feature vectors into corresponding embedding vectors, forming an initial embedding vector. The specific execution of the embedding mapping process is as follows: Numerical transformation is performed on each feature dimension of the high-dimensional diagnostic feature vector, mapping the numerical values ​​of each feature dimension to the corresponding embedding values. Dimensional compression is then performed on the numerically mapped embedding values ​​to map them to the corresponding embedding dimension range. The compressed embedding values ​​are then arranged in order to form a continuous vector structure, resulting in an embedding vector that corresponds one-to-one with the original high-dimensional diagnostic feature vector, which serves as the initial embedding vector. Based on the initial embedding vectors and the diagnostic labels corresponding to each medical image sample, the initial embedding vectors belonging to the same diagnostic category are aggregated to obtain the category representation of each diagnostic category. Furthermore, a sample constraint relationship is constructed where samples of the same category are close to each other, and samples of different categories are separated from each other. The diagnostic labels corresponding to each medical image sample are derived from the diagnostic category labels contained in each data point of the standardized medical image data. The aggregation process for initial embedding vectors belonging to the same diagnostic category specifically involves: Initial embedding vectors with the same diagnostic label are grouped into the same category set. Numerical aggregation processing is performed on each embedding vector in the same category set according to the corresponding dimension. The values ​​in the same dimension are averaged to generate the aggregated vector of the corresponding category. The aggregated vector is used as the category representation of the corresponding diagnostic category. Based on the sample constraint relationship, clustering consistency optimization is performed on the initial embedding vector in the embedding space. The position of each embedding vector in the embedding space is adjusted, the distance between embedding vectors belonging to the same diagnostic category is gradually reduced, and the distance between embedding vectors belonging to different diagnostic categories is gradually increased, so as to obtain the optimized embedding vector. The optimized embedding vectors are subjected to spatial structure mapping processing. Based on the distribution structure of the optimized embedding vectors in the embedding space, a new diagnostic feature representation is formed, resulting in a diagnostic feature vector processed by spatial structure mapping. Specifically, the spatial structure mapping processing of the optimized embedding vectors involves: Based on the positional relationship of the optimized embedding vectors in the embedding space, spatial distribution analysis is performed on each embedding vector. The embedding vectors are grouped according to their relative distance, adjacency relationship and arrangement structure in space. Dimensional combination and positional encoding processing are performed on each group of embedding vectors according to the group. The dimension value of each embedding vector is numerically bound to its position index in the embedding space. The bound embedding values ​​are then matrix-reorganized to form a new diagnostic feature representation, resulting in a diagnostic feature vector processed by spatial structure mapping.

[0025] In this embodiment, the calculation of the correlation metric between each feature dimension and the diagnostic label, and the redundancy metric between each feature dimension, includes: The values ​​of each medical image sample in each feature dimension of the diagnostic feature vector processed by spatial structure mapping are summarized to form the corresponding feature dimension sequence, and the known diagnostic category information of each medical image sample is obtained to form a diagnostic category sequence. For each feature dimension, based on the joint distribution relationship between the feature dimension sequence and the diagnostic category sequence, a correlation measure between the feature dimension and the diagnostic category is calculated. Specifically, the calculation of the correlation measure between the feature dimension and the diagnostic category involves: The range, frequency, and trend of feature dimensions in different diagnostic categories are statistically analyzed, and the statistical results are jointly grouped with the corresponding diagnostic category labels. The range, frequency, and trend of values ​​belonging to the same diagnostic category are cumulatively processed to obtain cumulative statistical values. The range, frequency, and trend of values ​​of values ​​in different diagnostic categories are differentially processed to obtain comparative statistical values. The numerical difference between the cumulative statistical values ​​and the comparative statistical values ​​is used to perform a measurement of the correlation between the corresponding feature dimensions and the diagnostic categories. For any two different feature dimensions, based on the changing trend relationship between the corresponding feature dimension sequences, a redundancy measure between the two feature dimensions is calculated. Specifically, the calculation of the redundancy measure between the two feature dimensions involves: The changes in the values ​​of the two feature dimension sequences in each medical image sample are compared sample by sample. The differences in values ​​at corresponding positions and the direction of value change are recorded synchronously. The samples with the same direction of change of the two feature dimension sequences are cumulatively statistically analyzed, and the samples with small differences are cumulatively statistically analyzed. The statistical values ​​of the consistent direction and the cumulative statistical values ​​of the difference are weighted and combined to form a redundancy measure value.

[0026] In this embodiment, obtaining the optimal diagnostic feature subset includes: Based on correlation and redundancy measures, and combined with the diagnostic feature vector processed by spatial structure mapping, the feature dimensions in the diagnostic feature vector are collected to construct a set of original feature dimensions to be screened, and a corresponding importance record entry is established for each original feature dimension. The Boruta algorithm is executed. For each original feature dimension in the set of original feature dimensions to be filtered, a structure-preserving perturbation mechanism is introduced. This perturbs the value distribution of the original feature dimension while preserving the local structural relationships, generating a corresponding shadow feature dimension. The number of shadow feature dimensions is consistent with the number of original feature dimensions, forming a shadow feature set, where: The structure-preserving perturbation mechanism refers to perturbing the value distribution of the original feature dimension by only making slight random adjustments to the numerical value of each value, without changing the value ordering and adjacent position relationship of the original feature dimension in the same diagnostic category sequence. By adding perturbation increments of different magnitudes to each value, the perturbed feature dimension changes numerically but remains consistent in local ordering and adjacent value interval relationship, forming a structure-preserving perturbation value sequence. The generation of a corresponding shadow feature dimension is specifically as follows: Based on the perturbation value sequence after structure preservation perturbation, the perturbation value sequence replaces the values ​​of each sample position in the original feature dimension item by item according to the sample arrangement order of the original feature dimension, forming a feature dimension sequence with the same value arrangement structure as the original feature dimension but whose values ​​have been perturbed and adjusted, which serves as the shadow feature dimension; The formation of the shadow feature set specifically includes: For each of the original feature dimensions to be screened, perform structure preservation perturbation and shadow feature dimension generation operations respectively. Generate a corresponding shadow feature dimension for each original feature dimension. Collect and organize all the generated shadow feature dimensions according to the index order of the original feature dimensions, and store them uniformly after numbering them according to the feature dimension sequence to form a shadow feature set with the same number of original feature dimensions. For each original feature dimension, a category discrimination score is calculated based on the differences in value distribution across different diagnostic categories. Similarly, for each shadow feature dimension in the shadow feature set, the corresponding shadow category discrimination score is calculated in the same manner, where: The calculation of the category discrimination score is as follows: Statistical processing is performed on the value sequences of the original feature dimensions in each diagnostic category. The mean value, frequency distribution of the value, and statistical value of the direction of value change are calculated for each diagnostic category. The statistical results corresponding to different diagnostic categories are compared by difference. The difference of mean, frequency, and direction of change are accumulated by numerical processing. The accumulated values ​​are then weighted and summed to obtain the category discrimination score corresponding to the original feature dimension. The calculation of the corresponding shadow category discrimination score is as follows: For each shadow feature dimension in the shadow feature set, perform value statistical processing according to the same diagnostic category grouping method as the original feature dimension to obtain the shadow category discrimination score of the shadow feature dimension; For each original feature dimension, the intra-class consistency change is calculated based on the value fluctuations within each diagnostic category. The corresponding intra-class consistency change is then calculated for each shadow feature dimension in the shadow feature set in the same manner. The calculation of the intra-class consistency change is as follows: For the value sequence of the original feature dimension in the same diagnostic category, fluctuation statistics processing is performed. The cumulative value of the change amplitude, the cumulative value of the fluctuation frequency, and the consistency ratio of the change direction of each sample value in the corresponding diagnostic category are calculated respectively. The cumulative value of the change amplitude, the cumulative value of the fluctuation frequency, and the consistency ratio of the change direction are weighted and combined. The weighted combination result is used as the intra-class consistency change amount in the corresponding diagnostic category. The intra-class consistency change amount in all diagnostic categories is accumulated to obtain the intra-class consistency change amount of the corresponding original feature dimension. The calculation of the corresponding consistency change within the shadow class is as follows: For each shadow feature dimension in the shadow feature set, perform value fluctuation statistical processing according to the same diagnostic category as the original feature dimension to obtain the corresponding shadow class consistency change; The importance statistics of the original features are constructed based on the category discrimination score and intra-class consistency change of the original feature dimension. The importance benchmark of the shadow features is constructed based on the corresponding statistics of the shadow feature dimension. A dynamic threshold elimination mechanism is introduced to retain or eliminate features based on whether the importance statistics of the original features are significantly higher than the importance benchmark of the shadow features. The construction of the importance statistics of the original features is as follows: For each original feature dimension, the corresponding category discrimination score and intra-class consistency change are weighted and combined. The weighted combination value is then normalized and mapped to a comparable value under a uniform scale. The normalized value is then used as the importance statistic corresponding to the current original feature dimension. The construction of the importance benchmark for the shadow features is as follows: For each shadow feature dimension in the shadow feature set, the corresponding shadow category discrimination score and the shadow intra-class consistency change are weighted and combined according to the same weighting ratio as the original feature. The weighted combination values ​​are normalized and mapped. The normalized values ​​corresponding to all shadow feature dimensions are aggregated and statistically analyzed to calculate the median, upper quantile, and lower quantile values, which serve as the importance benchmarks for shadow features. The dynamic threshold elimination mechanism refers to dynamically generating a threshold interval based on the median and quantile range of the shadow feature's importance benchmark when comparing the importance statistics of the original feature with the importance benchmark of the shadow feature. The threshold interval is used as the boundary for judging feature retention and elimination. When the importance statistics of the original feature are higher than the upper limit of the threshold interval, it is marked as retained; when they are lower than the lower limit of the threshold interval, it is marked as eliminated; when they are between the threshold intervals, it is marked as pending and will continue to be evaluated and processed in the next iteration. In each iteration, the corresponding shadow feature dimensions are regenerated for the retained original feature dimensions, and the calculation of class discrimination scores, intra-class consistency changes, and dynamic removal are repeated. An adaptive termination strategy is introduced, determining whether to terminate the iteration based on whether the change in the number of retained original features has reached a stable state. The adaptive termination strategy refers to recording the number of original feature dimensions retained in the current iteration after each iteration and comparing it with the number retained in the previous iteration. When the difference in the number of features retained in multiple consecutive iterations is less than the change threshold, the current iteration is marked as the termination iteration. After the iteration terminates, the set of original feature dimensions that are retained is taken as the optimal diagnostic feature subset.

[0027] In this embodiment, generating the diagnostic result includes: Based on the optimal diagnostic feature subset, a corresponding feature dimension index set is determined, and a correspondence is established between the feature dimension index set and the diagnostic feature vector processed by spatial structure mapping. Specifically, determining the corresponding feature dimension index set involves: The optimal subset of diagnostic features retained after iterative filtering is parsed dimension by dimension. The position number of each feature dimension in the diagnostic feature vector after spatial structure mapping is recorded. All the position numbers of the dimension are arranged in order of appearance of the feature dimension to form the corresponding feature dimension index set. Based on the feature dimension index set, the values ​​of the corresponding feature dimensions are extracted from the diagnostic feature vector corresponding to each medical image sample after spatial structure mapping, forming the corresponding feature sub-vector to be diagnosed. Based on the diagnostic feature subvectors of medical image samples and the corresponding known diagnostic category information, the feature distribution of the diagnostic feature subvectors is statistically analyzed for each diagnostic category to form a reference feature distribution set for each diagnostic category. The diagnostic feature subvectors of the medical image samples to be diagnosed are compared with the reference feature distribution set. A distribution difference metric is calculated between the vectors and the reference feature distributions for each diagnostic category, and a corresponding category matching degree set is generated, where: The distribution comparison process is performed as follows: The values ​​of the diagnostic feature sub-vectors of the medical image samples to be diagnosed are statistically processed in each feature dimension to form a feature distribution sequence of the diagnostic feature sub-vectors. The feature distribution sequence is then aligned and compared with the reference feature distributions corresponding to each diagnostic category in the reference feature distribution set. The alignment and comparison are performed dimension by dimension according to the feature dimension index. The values ​​of the corresponding feature dimension are compared with the statistical range of the corresponding dimension in the reference distribution to generate a dimension-by-dimensional comparison record. The calculation of the distribution difference measure between the distribution and the reference feature distribution of each diagnostic category is specifically as follows: For each feature dimension in the dimension-by-dimensional comparison record, the deviation value, excess ratio and direction of change between the value of the feature dimension and the reference statistical range are calculated and weighted and accumulated. The accumulated difference index is used as the difference score corresponding to the feature dimension. The difference scores of all feature dimensions are merged and normalized to obtain the distribution difference measurement value corresponding to the diagnostic category. The generation of the corresponding category matching degree set is specifically as follows: The distribution difference measurement values ​​corresponding to each diagnostic category are reversed and mapped. Difference measurements below the difference threshold are mapped to matching values ​​above the matching benchmark, and difference measurements above the difference threshold are mapped to matching values ​​below the matching benchmark. The mapped matching values ​​are then uniformly scaled and adjusted to generate matching values ​​for each diagnostic category. The matching values ​​corresponding to all diagnostic categories are then combined in order of diagnostic category to form a category matching set. The target diagnostic category corresponding to the medical image sample to be diagnosed is determined based on the category matching degree set, and the target diagnostic category is output as the diagnostic result. Specifically, the determination of the target diagnostic category corresponding to the medical image sample to be diagnosed based on the category matching degree set is as follows: The matching values ​​corresponding to each diagnostic category in the category matching set are compared item by item. The maximum matching value is found and the diagnostic category corresponding to the maximum matching value is marked as the target diagnostic category of the medical image sample to be diagnosed.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to the early diagnosis project of pulmonary nodules in the chest imaging screening center of a tertiary hospital. This hospital processes approximately 1200 chest CT scans daily, with about 82% of patients being routine health checkups. These patients exhibit significant differences in imaging characteristics, with small nodules, ground-glass opacities, and atypical shadows being the most easily misinterpreted types. Because early pulmonary nodules have weak imaging signals, indistinct borders, and significant size variations, manual interpretation is easily limited by interpreter fatigue and subjective experience, especially for tiny lesions smaller than 5mm, where the accuracy rate is low. Statistics from 2024 show that the hospital's misdiagnosis rate for small nodules was approximately 14.6%, and the missed diagnosis rate was approximately 11.2%, failing to meet the requirements for early diagnosis.

[0029] In the diagnostic project using the method proposed in this invention, the hospital adopted the computer-aided diagnostic method based on deep learning. First, the acquired chest CT medical images were standardized and preprocessed. Then, an improved MoCo model was used to represent the features of the images. Heterogeneous query coding and momentum coding strategies were employed to jointly represent the local texture of lesions and the overall structural features of the lungs, generating high-dimensional feature vectors with good diagnostic discriminative power. Subsequently, an improved Boruta algorithm was used to select a stable subset of features reflecting lesion diagnostic differences from the high-dimensional features. Based on the selected feature distribution and category matching results, an auxiliary diagnostic category was output. The entire process can complete feature representation and preliminary diagnosis without additional manual annotation.

[0030] The hospital selected 47,620 chest CT images collected between January and August 2025 as test samples, with each image marked as independently reviewed by five associate chief physicians and cross-validated by multiple individuals. The method of this invention achieved an overall diagnostic accuracy of 92.8% in pulmonary nodule screening, with an accuracy rate of 89.5% for identifying ground-glass nodules. Particularly in cases with nodule diameters less than 5 mm, the accuracy rate increased from 78.2% with traditional screening systems to 88.7%, while the misdiagnosis and missed diagnosis rates decreased to 6.3% and 5.0%, respectively. In approximately 3,800 complex cases, the average image interpretation time for doctors using the system was reduced from 7 minutes per case to 4.2 minutes.

[0031] This embodiment shows that the present invention can improve the accuracy of identifying complex imaging lesions without the need for extensive manual annotation, effectively reduce the risk of misdiagnosis and missed diagnosis, and significantly improve the diagnostic efficiency of chest CT, thus having practical clinical application value.

[0032] Table 1. Performance comparison of the method of the present invention with that of traditional auxiliary diagnostic systems in CT screening of pulmonary nodules.

[0033] As can be seen from the comparative data in Table 1, the deep learning-based computer-aided diagnostic method proposed in this invention has significant performance advantages in clinical pulmonary nodule CT image screening. In terms of overall diagnostic capability, without relying on extensive manual annotation, the method of this invention improves the overall diagnostic accuracy from 85.3% of the traditional system to 92.8%, sensitivity from 84.1% to 91.6%, and specificity from 86.8% to 93.1%. This indicates that the medical image features extracted and screened by the method of this invention can maintain high disease identification and discriminative power under unsupervised conditions, solving the problems of strong dependence on annotation and insufficient model generalization in existing schemes.

[0034] The comparison of recognition results for different lesion types shows that this invention demonstrates significant advantages in recognizing both typical nodules and complex lesions. Specifically, the recognition rate for typical solid nodules increased from 88.9% to 94.2%, and the recognition rate for ground-glass opacities increased dramatically from 76.4% to 89.5%, with the recognition rate for micronodules particularly improving from 78.2% to 88.7%. This indicates that the heterogeneous contrastive learning feature extraction structure proposed in this invention can more accurately represent subtle textures and local structural features, solving the problem of unstable recognition of early-signal-weakened lesions in traditional diagnostic methods.

[0035] In terms of reducing the risk of misdiagnosis and missed diagnosis, the method of this invention reduces the misdiagnosis rate from 10.5% to 6.3% and the missed diagnosis rate from 11.2% to 5.0%. The feature selection process based on the improved Boruta algorithm makes the selected feature subset more stable and diagnostically distinct, effectively filtering out noise and redundant information, making the output diagnostic results more reliable, thereby reducing the interference of low-quality features on the recognition results.

[0036] In terms of clinical efficiency, the success rate of the method described in this invention for assisting in the screening of difficult cases reaches 86.8%, an improvement of 15.2% compared to traditional methods. This effectively helps doctors quickly locate potentially suspicious areas and provides highly reliable auxiliary references. In large-scale image review scenarios, the average image review time for doctors using the system described in this invention is reduced from 7.0 minutes per case to 4.2 minutes, decreasing the initial interpretation time by approximately 40%, thus improving clinical work efficiency and reducing the workload of doctors.

[0037] As can be seen from the above data, the method of the present invention has significant advantages in extracting deep diagnostic features, improving the accuracy of image recognition, reducing the risk of misdiagnosis and missed diagnosis, and improving the efficiency of clinical screening. It can effectively make up for the shortcomings of traditional diagnostic methods in complex image analysis and has high clinical promotion value.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A computer-aided diagnostic method based on deep learning, characterized in that, include: Collect medical imaging data, preprocess the medical imaging data, and obtain standardized medical imaging data; Based on standardized medical image data, an improved MoCo model is constructed. The first encoder and the second encoder of the improved MoCo model adopt a heterogeneous network structure to obtain a high-dimensional diagnostic feature vector. Based on the high-dimensional diagnostic feature vector, embedding space clustering consistency optimization is performed to minimize the embedding distance of samples of the same class and maximize the embedding distance of samples of different classes. Spatial structure mapping is then performed on the high-dimensional diagnostic feature vector to obtain the diagnostic feature vector processed by spatial structure mapping. Based on the diagnostic feature vector processed by spatial structure mapping, the correlation measure between each feature dimension and the diagnostic label and the redundancy measure between each feature dimension are calculated. Based on correlation and redundancy measures, the Boruta algorithm is used to construct a shadow feature set with the same number of original feature dimensions. The importance of the original feature dimensions and the shadow feature dimensions is statistically tested, and unimportant feature dimensions are iteratively eliminated to obtain the optimal diagnostic feature subset. Based on the optimal diagnostic feature subset, the corresponding feature dimensions are extracted from the diagnostic feature vectors processed by spatial structure mapping to form the feature subvectors to be diagnosed, and feature distribution comparison analysis is performed to generate diagnostic results.

2. The computer-aided diagnostic method based on deep learning according to claim 1, characterized in that, The medical imaging data specifically includes X-ray imaging data, CT imaging data, MRI imaging data, nuclear medicine imaging data, and ultrasound imaging data.

3. The computer-aided diagnostic method based on deep learning according to claim 1, characterized in that, The preprocessing of medical image data specifically includes grayscale normalization, noise suppression and structure-preserving filtering, artifact correction, automatic extraction of regions of interest, spatial geometric correction, and resolution resampling.

4. The computer-aided diagnostic method based on deep learning according to claim 1, characterized in that, The obtained high-dimensional diagnostic feature vector includes: An improved MoCo model is constructed, which consists of a heterogeneous query encoder module, a heterogeneous momentum encoder module, a dynamic momentum update module, a semantic alignment feature queue module, and a contrastive loss calculation module. Based on standardized medical image data, a first data augmentation operation and a second data augmentation operation are performed respectively to form a first view and a second view; The first view is input into the heterogeneous query encoder module, and variable convolution kernel processing and dual-path downsampling processing are performed on the first view to form downsampling features at different scales. Local stability enhancement feature fusion processing is then performed, and a differentiated fusion strategy is adopted for smooth and variable regions to obtain query feature representations. The second view is input into the heterogeneous momentum encoder module for adaptive patch division. The patch size is adjusted according to the structural complexity of the image region. The patch sequence is converted into a semantic representation. An attention guidance mechanism is introduced to adaptively adjust the attention weight allocation according to the gray density difference of different regions. Then, a feedforward transformation process for global consistency reconstruction is performed to obtain the key feature representation by reconstructing the structural relationship between patches. The parameters of the heterogeneous query encoder module are obtained by the dynamic momentum update module as the target parameters. A momentum adjustment strategy is introduced to adaptively adjust the momentum update coefficient based on the difference between the parameters of the heterogeneous momentum encoder module and the target parameters. The current update amplitude is determined based on the momentum update coefficient, and the parameters of the heterogeneous momentum encoder module are adjusted to gradually approach the target parameters. Input the key feature representation into the semantic alignment feature queue module, delete the earliest key feature representation that entered the queue, and perform replacement processing based on the semantic difference between the key feature representation and the key feature representation in the queue to obtain the updated semantic alignment feature queue. The contrastive loss calculation module performs contrastive learning operations based on the query feature representation, key feature representation, and updated semantic alignment feature queue, and introduces a semantic interval control mechanism to dynamically adjust the contrastive learning scale to generate a high-dimensional diagnostic feature vector.

5. The computer-aided diagnostic method based on deep learning according to claim 1, characterized in that, The obtained diagnostic feature vector after spatial structure mapping includes: Embedding mapping is performed on the high-dimensional diagnostic feature vectors of each medical image sample to convert the high-dimensional diagnostic feature vectors into corresponding embedding vectors, forming the initial embedding vectors. Based on the initial embedding vector and the diagnostic labels corresponding to each medical image sample, the initial embedding vectors belonging to the same diagnostic category are aggregated to obtain the category representation of each diagnostic category, and a sample constraint relationship is constructed in which samples of the same category are close to each other and samples of different categories are separated from each other. Based on the sample constraint relationship, clustering consistency optimization is performed on the initial embedding vector in the embedding space. The position of each embedding vector in the embedding space is adjusted, the distance between embedding vectors belonging to the same diagnostic category is gradually reduced, and the distance between embedding vectors belonging to different diagnostic categories is gradually increased, so as to obtain the optimized embedding vector. The optimized embedding vectors are subjected to spatial structure mapping processing. Based on the distribution structure of the optimized embedding vectors in the embedding space, a new diagnostic feature representation is formed, resulting in a diagnostic feature vector processed by spatial structure mapping.

6. The computer-aided diagnostic method based on deep learning according to claim 1, characterized in that, The calculation of the correlation metric between each feature dimension and the diagnostic label, and the redundancy metric between each feature dimension, includes: The values ​​of each medical image sample in each feature dimension of the diagnostic feature vector processed by spatial structure mapping are summarized to form the corresponding feature dimension sequence, and the known diagnostic category information of each medical image sample is obtained to form a diagnostic category sequence. For each feature dimension, a correlation measure between the feature dimension and the diagnostic category is calculated based on the joint distribution relationship between the feature dimension sequence and the diagnostic category sequence. For any two different feature dimensions, calculate the redundancy measure between the two feature dimensions based on the changing trend relationship between the corresponding feature dimension sequences.

7. The computer-aided diagnostic method based on deep learning according to claim 1, characterized in that, The process of obtaining the optimal diagnostic feature subset includes: Based on correlation and redundancy measures, and combined with the diagnostic feature vector processed by spatial structure mapping, the feature dimensions in the diagnostic feature vector are collected to construct a set of original feature dimensions to be screened, and a corresponding importance record entry is established for each original feature dimension. The Boruta algorithm is executed. For each original feature dimension in the set of original feature dimensions to be screened, a structure-preserving perturbation mechanism is introduced to perturb the value distribution of the original feature dimension while keeping the local structural relationship unchanged, generating a corresponding shadow feature dimension. The number of shadow feature dimensions is consistent with the number of original feature dimensions, forming a shadow feature set. For each original feature dimension, the category discrimination score is calculated based on the differences in the value distribution in different diagnostic categories, and the corresponding shadow category discrimination score is calculated for each shadow feature dimension in the shadow feature set in the same way; For each original feature dimension, the intra-class consistency change is calculated based on the value fluctuation within each diagnostic category, and the corresponding intra-class consistency change is calculated for each shadow feature dimension in the shadow feature set in the same way. The importance statistics of the original features are constructed based on the category discrimination score and intra-class consistency change of the original feature dimension. The importance benchmark of the shadow feature is constructed based on the corresponding statistics of the shadow feature dimension. A dynamic threshold elimination mechanism is introduced to retain or eliminate features based on whether the importance statistics of the original features are significantly higher than the importance benchmark of the shadow features. In each iteration, the corresponding shadow feature dimension is regenerated for the retained original feature dimension, and the calculation of class discrimination score, intra-class consistency change amount and dynamic elimination are repeatedly performed. An adaptive termination strategy is introduced to determine whether to terminate the iteration based on whether the change in the number of retained original features has reached a stable state. After the iteration terminates, the set of original feature dimensions that are retained is taken as the optimal diagnostic feature subset.

8. The computer-aided diagnostic method based on deep learning according to claim 1, characterized in that, The generation of diagnostic results includes: Based on the optimal diagnostic feature subset, the corresponding feature dimension index set is determined, and a correspondence is established between the feature dimension index set and the diagnostic feature vector processed by spatial structure mapping. Based on the feature dimension index set, the values ​​of the corresponding feature dimensions are extracted from the diagnostic feature vector corresponding to each medical image sample after spatial structure mapping, forming the corresponding feature sub-vector to be diagnosed. Based on the diagnostic feature subvectors of medical image samples and the corresponding known diagnostic category information, the feature distribution of the diagnostic feature subvectors is statistically analyzed for each diagnostic category to form a reference feature distribution set for each diagnostic category. The diagnostic feature subvectors of the medical image samples to be diagnosed are compared with the reference feature distribution set. The distribution difference measure between the vectors and the reference feature distributions of each diagnostic category is calculated, and the corresponding category matching degree set is generated. The target diagnostic category corresponding to the medical image sample to be diagnosed is determined based on the category matching degree set, and the target diagnostic category is output as the diagnostic result.

Images