Pathological space diagram modeling system and method for intelligent diagnosis of skin lesion

Abstract

The invention belongs to the technical field of dermatoscope image analysis, and relates to a pathological space diagram modeling system and method for skin lesion intelligent diagnosis, the system comprises four parallel deployment visual representation learning and topological structure modeling, the modeling comprises a saliency map embedding layer and a plurality of pathological space diagram blocks, the pathological space graph block is sequentially composed of a pathological feature point digger, a layered graph connector and a geometric distribution learner; the method comprises the following steps: constructing a multi-scale saliency block embedding layer to extract initial visual representation, introducing a pathological feature point digger to realize intelligent identification of sparse key pathological points, constructing a hierarchical graph connector to organize a multi-scale topological relation, realizing hierarchical topological reasoning and vector distillation of a geometric distribution learner, and constructing a geometric distribution learner. Defining an attention fusion mechanism and a classification head to realize topological guidance diagnosis; according to the method, the defects in the aspects of structural modeling, calculation distribution, graph representation and the like in the prior art are overcome, and the diagnosis accuracy, robustness and clinical interpretability are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of dermoscopic image analysis technology, specifically relating to a pathological spatial map modeling system and method for intelligent diagnosis of skin lesions. Background Technology

[0002] In the field of skin lesion diagnosis, current technologies mainly rely on deep learning algorithms to achieve automated analysis. Convolutional Neural Network (CNN)-based methods have made significant progress. To further enhance feature representation capabilities, some methods employ attention mechanisms for optimization. Furthermore, ensemble learning methods such as CFI-Net, which uses Choquet fuzzy integrals to fuse predictions from multiple base models, and HGRE, which utilizes hyperbolic space to model hierarchical relationships between lesion categories, have all improved diagnostic accuracy to varying degrees.

[0003] Recently, graph neural networks (GNNs) have been explored for dermoscopy image analysis to explicitly model non-Euclidean topological relationships between pathological features. One class of methods uses superpixels as graph nodes, while another directly uses regular image patches as nodes to construct pseudo-graph structures. These methods have made initial explorations into the potential of structured representation, but the graph construction strategies are overly simplistic and fail to fully reflect the key pathological units (such as pigment networks, stripes, and spheres) and their multi-scale spatial layouts that are of interest in clinical diagnosis.

[0004] However, existing dermoscopy image analysis methods still have the following shortcomings: (1) Lack of explicit modeling ability for pathological topology. Existing deep learning methods generally treat dermoscopic images as a regular grid of pixels or image blocks. Whether it is the local receptive field of CNN or the global attention mechanism of ViT, it is difficult to accurately capture pathological features with clear clinical significance (such as blue-white curtain, dotted structures, stripes) and their complex spatial arrangement relationships. This "black box" feature learning method cannot simulate the topological analysis process when dermatologists rely on ABCDE rules for diagnosis, resulting in the model's inability to effectively utilize the spatial grammatical information between structures when distinguishing between malignant melanoma and benign lesions.

[0005] (2) Lack of clinical prior guidance in the allocation of computational resources. The attention mechanism of traditional Transformer is semantically blind, treating the relationships between all image patches equally, lacking structured cognition based on clinical pathology knowledge. This may cause attention to be diluted by irrelevant background textures, making it difficult to focus on key pathological units and their interactions. Even though state-space models such as Mamba have improved computational efficiency, their serialization processing method still does not solve the problem of how to explicitly construct a graph structure composed of irregular, semantically clear pathological primitives.

[0006] (3) Mismatch between graph construction strategy and pathological structure. Existing graph neural networks (GNNs) use superpixels or regular image patches as nodes. The former often fails to accurately align its boundaries with complete pathological structures with single clinical significance, resulting in the loss of fine-grained structural information; the latter essentially reproduces the local connectivity of CNNs, failing to leverage the advantages of GNNs in handling arbitrary connections. This limitation of graph construction strategy hinders the model from learning the multi-scale, hierarchical topological relationships of lesions.

[0007] Therefore, there is a need for a pathological spatial map modeling method for intelligent diagnosis of skin lesions that can improve diagnostic accuracy, robustness, and clinical interpretability, and solve the above-mentioned technical problems. Summary of the Invention

[0008] This invention provides the following technical solution: a pathological spatial map modeling system for intelligent diagnosis of skin lesions, comprising: A pathological feature point miner is used to generate pathological correlation maps through a lightweight scoring network, and a deterministic Top-K sampling strategy is used to select a fixed number of K pathological feature points with the highest diagnostic value; a hierarchical graph connector is used to construct a composite graph structure; and a geometric distribution learner is used to perform multi-scale information propagation on the hierarchical graph.

[0009] The pathological spatial map modeling system includes four parallel deployments of visual representation learning and topology modeling. The parallel deployments of visual representation learning and topology modeling include a saliency map embedding layer and multiple pathological spatial map tiles. The pathological spatial map tiles are composed of a pathological feature point miner, a hierarchical map connector, and a geometric distribution learner in sequence. The pathological spatial map tiles transform lesion images into clinically inspired structured representations.

[0010] Parallel deployment of visual representation learning and topology modeling uses an attention fusion mechanism to iteratively inject the learned spatial distribution vectors into the main visual stream, so that the depth visual features are continuously guided by topological understanding, achieving synergistic enhancement of the two types of information.

[0011] Preferably, the pathological feature point miner concatenates heterogeneous features extracted from deep semantic feature maps, CIELab color space transformation, and Gabor filter banks for each pathological feature point to construct a rich composite feature vector; it concentrates computational resources on clinically significant structures, reducing the complexity of graph construction while retaining key diagnostic information.

[0012] When constructing a composite graph structure using a layered graph connector, the fine-grained graph uses PCP as nodes and employs a four-quadrant minimum distance connection algorithm to establish edges—for each source node, the nearest neighbor node in the coordinate quadrant is found based on the Manhattan distance; the coarse-grained graph generates region nodes through superpixel segmentation and defines edges based on spatial adjacency.

[0013] The geometric distribution learner first applies an N-layer graph attention network to learn the micro-interaction patterns between pathological feature points on the fine-grained graph; then, it aggregates the enhanced pathological feature point features to the corresponding region nodes through pooling operations, updating the node representation of the coarse-grained graph; next, it applies an M-layer GAT to the coarse-grained graph to model the macro-spatial distribution pattern; finally, it generates a compact spatial distribution vector through a global graph readout mechanism.

[0014] More preferably, the pathological spatial map modeling system adopts a multi-stage decomposition framework: the first stage extracts initial visual features through a saliency map embedding layer; the second to fourth stages gradually construct a hierarchical map of pathological feature points and learn their spatial distribution patterns through cascaded pathological spatial map tiles; the topological vectors generated in each stage feed back into the visual feature flow through an attention fusion mechanism to achieve feature enhancement under structure guidance.

[0015] This invention also discloses a pathological spatial map modeling method for intelligent diagnosis of skin lesions. This modeling method is used in the aforementioned modeling system and includes the following steps: Step 1: Construct a multi-scale saliency block embedding layer to extract the initial visual representation; Step 2: Introduce a pathological feature point mining tool to achieve intelligent identification of sparse key pathological points; Step 3: Construct hierarchical graph connectors to organize multi-scale topological relationships; Step 4: Implement hierarchical topological reasoning and vector distillation for the geometric distribution learner; Step 5: Define the attention fusion mechanism and classification head to implement topology-guided diagnosis.

[0016] Preferably, step one specifically includes: The initial embedding layer of each stage receives the input image or the feature map from the previous stage and generates a salient feature map. The first stage uses a multi-convolution kernel strategy to capture local texture, mesoscopic structure and global lesion morphology respectively, and the output features are fused by concatenation. Subsequent stages use smaller kernels to refine the features.

[0017] The multi-scale feature fusion formula is as follows:

[0018] in, For the input image or feature map, Indicates the kernel size as Convolution operation, This indicates a splicing operation at the channel dimension.

[0019] Preferably, step two specifically includes: The dense feature representation is transformed into a sparse, semantically rich set of key points through a pathological feature point mining tool; the importance scoring network generates a pathological correlation map through convolutional layers and GELU / Sigmoid activation function to quantify the diagnostic value of each spatial location; the highest score point is selected by a Top-K sampling strategy, and its coordinates constitute a set of nodes.

[0020] For each node, features are extracted from three modalities: deep semantic features are directly indexed from the corresponding position in the feature map, color features convert RGB values ​​to CIELab space, and texture features extract local neighborhood patterns through Gabor filter banks; the three types of features are concatenated to form a composite node feature with dimension D_feat.

[0021] The specific calculation process of the importance scoring network is as follows:

[0022] in, For saliency feature maps, and For two different Convolutional layer This represents the GELU activation function. This is the Sigmoid function.

[0023] The formula for extracting heterogeneous features of PCP nodes is:

[0024] in For deep semantic features, This represents a color transformation function. This represents the texture extraction function of the Gabor filter bank. For nodes The local neighborhood, Indicates feature splicing, This represents the input image for the k-th image patch.

[0025] Preferably, step three specifically includes: Fine-grained graph construction employs a four-quadrant minimum distance connection algorithm: for each source node, the plane is divided into four quadrants centered on its coordinates, and the nearest neighbor node is found in each quadrant based on the Manhattan distance to establish a bidirectional edge.

[0026] Meanwhile, the coarse-grained graph generates region nodes by superpixel segmentation of image blocks and defines edge sets based on spatial adjacency; the mapping function assigns each PCP to the corresponding region node to establish cross-level associations; the final output hierarchical graph G_h is a pair of tuples, which simultaneously supports micro and macro topological analysis.

[0027] The formula for constructing the edge set of the minimum distance connection in the four quadrants is:

[0028]

[0029] in Relative to node The The set of nodes in each quadrant For Manhattan distance, The value of the variable that makes the function reach its minimum value. Represents the k-th feature node or vertex. Represents the set of all selected fine-grained feature nodes. Represents the x-coordinate of the i-th feature node in the image. Represents the x-coordinate of the k-th feature node in the image. Represents the ordinate of the i-th feature node in the image, This represents the ordinate of the k-th feature node in the image.

[0030] Preferably, step four specifically includes: The geometrically distributed learner performs a bottom-up inference process; firstly, an N=4-layer graph attention network is applied to the fine-grained graph to dynamically update the features of pathological nodes and learn the contextual interactions between local pathological points; the attention weights of the GAT layer are calculated as follows:

[0031] in, For nodes Features The weight matrix is ​​a learnable matrix. For attention weight vectors, This indicates a splicing operation. For nodes The neighborhood group, For activation function, The feature vector of node j The feature vector of node k The transpose operation represents a matrix or vector. This represents the natural exponential function.

[0032] Subsequently, the enhanced PCP features are aggregated to the corresponding region nodes through pooling operations, generating the region-enhanced features as follows:

[0033] in, This indicates the aggregation of a set of node features. This function represents mapping node vi to its region. The identifier of the j-th region or supernode This represents the updated feature vector of the i-th node at the current stage. Indicates the index of the node; AGGREGATE can use permutation-invariant functions such as mean, maximum, or summation.

[0034] Next, an M=3 layer GAT is applied to the coarse-grained map to model the macroscopic spatial distribution pattern; finally, a global graph readout mechanism is applied to perform global average pooling on the final region node features and input them into a multilayer perceptron to generate a spatial distribution vector D with dimension D_dist.

[0035] Preferably, step five specifically includes: First, it is expanded to include saliency feature maps. With the same spatial dimension, weights are then calculated using a channel attention mechanism:

[0036] in, Indicates a fully connected layer. It is the ReLU activation function. For the Sigmoid function, This represents the original feature vector or feature map of the i-th branch or stage; the fused features are:

[0037] in, This represents element-wise multiplication. This represents the deep feature map extracted by the backbone network; the process is executed iteratively in four stages, ensuring that the deep features are continuously guided by structural understanding.

[0038] After the final stage, the enhanced feature map and the set of all spatial distribution vectors are... The classification head is input together; the classification head consists of a global average pooling layer and a fully connected layer, and outputs the predicted probability of lesion type. Training uses cross-entropy loss:

[0039] in, For real labels, To predict probabilities, This represents the number of categories.

[0040] The beneficial effects of this invention are: 1. The DermPrism multi-stage network framework of the present invention realizes explicit decomposition and learning of pathological feature topology through pathological spatial tiles, systematically solves the shortcomings of existing technologies in structural modeling, computational allocation and graph representation, and significantly improves diagnostic accuracy, robustness and clinical interpretability.

[0041] 2. This invention reconstructs the diagnostic task into a visual-topological co-learning process using PSG blocks. A PCP miner identifies clinical salient points, a hierarchical graph connector constructs a multi-scale topological map, and a geometric distribution learner infers spatial distribution patterns. This design enables the model not only to identify "what" features but also to understand "where" and "how" they are organized. Experiments show that in the Derm7pt 7-point checklist assessment, the accuracy rates for identifying topologically sensitive features such as regression structures (RS) and stripes (STR) reach 83.8% and 77.8%, respectively, significantly outperforming existing methods. Therefore, this invention achieves explicit modeling of pathological topology, narrowing the logical gap with expert diagnosis.

[0042] 3. The PCP mining tool of this invention filters K=256 key pathological points through a learnable scoring network, reducing the computational complexity from O(HW) to O(K), ensuring that resources are concentrated on clinically significant structures such as blue-white curtains, small spheres, and stripes. This mechanism simulates the selective attention of experts during diagnosis, improving computational efficiency while enhancing feature discrimination; therefore, this invention can optimize the allocation of computational resources and focus on key diagnostic regions.

[0043] 4. This invention's hierarchical graph connector innovatively integrates fine-grained PCP graphs and coarse-grained region graphs: the fine-grained graph uses four-quadrant minimum distance connections to ensure a balance between global connectivity and sparsity; the coarse-grained graph models region adjacency relationships. This dual-hierarchical structure simultaneously supports joint reasoning of micro-texture and macro-layout, more closely aligning with the multi-scale diagnostic thinking of dermatologists. Comparative experiments show that the proposed graph construction strategy improves BAC by 1.34% compared to k-NN (k=8), demonstrating the superiority of the topology design; therefore, this invention can construct clinically aligned hierarchical graph representations, avoiding information granularity mismatch.

[0044] 5. This invention enhances knowledge transfer capabilities by explicitly learning the underlying structure of pathological topology, which is shared across datasets. Cross-dataset experiments show that, when trained on ISIC2020 and tested on Derm7pt, DermPrism achieves an average performance of 75.6%, a 1.8% improvement over the best baseline, with a particularly significant improvement of +6.4% on difficult-to-generalize point-like structures (DaG). Visualization confirms that the PCP miner can accurately locate high-risk structures, with the model focus highly consistent with clinical annotations, providing clear decision-making basis and significantly enhancing interpretability; therefore, this invention can enhance cross-dataset generalization ability and model interpretability.

[0045] 6. This invention achieves synergistic gains across modules through a multi-stage feature extraction, sparse keypoint sampling, hierarchical graph construction, topological inference, and attention fusion design. Ablation studies show that the complete framework improves BAC by +7.48% compared to the baseline, with the learnable PCP miner contributing the most (+3.91%), followed by the hierarchical structure and geometric distribution learner (+1.11% and +1.24%, respectively), validating the architecture's sophistication and necessity. This systematic design achieves state-of-the-art performance while maintaining a moderate number of parameters (approximately 25M), balancing accuracy and efficiency; therefore, this invention can achieve a performance leap through systematic synergistic optimization. Attached Figure Description

[0046] Figure 1 This is a diagram of the overall architecture of DermPrism, a pathological spatial map modeling system and method for intelligent diagnosis of skin lesions according to the present invention. Figure 2 This is a visualization comparing the identification results of the pathological feature point mining tool of the present invention with clinical annotations. Detailed Implementation

[0047] The related technologies of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0048] like Figures 1-2 As shown, this embodiment proposes DermPrism, a novel multi-stage network framework for skin lesion diagnosis. Its core lies in the decomposition and learning of pathological feature topology through pathological spatial graph (PSG) blocks, effectively overcoming problems such as insufficient structural modeling capabilities, lack of guidance in computational resource allocation, and graph representation mismatch in existing technologies. Compared with existing technologies, this invention has the following significant differences and improvements in overall architecture design and key module implementation: 1) Introducing PSG blocks to achieve visual-topology collaborative learning Existing methods treat diagnosis as a single visual feature extraction task or treat graph learning as a post-processing step, failing to achieve deep coupling. This invention deploys visual representation learning and topological modeling in parallel across four progressive stages. Each stage includes a saliency graph embedding layer and multiple PSG blocks. Each PSG block consists of a pathological feature point miner, a hierarchical graph connector, and a geometric distribution learner, transforming lesion images into clinically inspired structured representations. Through an attention fusion mechanism, the spatial distribution vectors learned in each stage are iteratively injected into the main visual flow, continuously guiding deep visual features with topological understanding and achieving synergistic enhancement of both types of information.

[0049] 2) Design a pathological feature point (PCP) mining tool to achieve sparse critical pathological point identification. Unlike computationally intensive paradigms that process all pixels, the pathological feature point miner generates pathological relevance maps through a lightweight scoring network and employs a deterministic Top-K sampling strategy to select a fixed number of K pathological feature points with the highest diagnostic value. For each pathological feature point, heterogeneous features extracted from the deep semantic feature map, CIELab color space transformation, and Gabor filter bank are concatenated to construct a rich composite feature vector. This design concentrates computational resources on clinically salient structures, significantly reducing the complexity of map construction while preserving key diagnostic information.

[0050] 3) A hierarchical graph connector is proposed to construct a multi-scale topological graph. To simultaneously capture microscopic textures and macroscopic layouts, a layered graph connector constructs a composite graph structure: the fine-grained graph uses PCPs as nodes and employs a four-quadrant minimum distance connection algorithm to establish edges—for each source node, the nearest neighbor node in the four quadrants of its coordinate system is found by Manhattan distance, ensuring a balance between global connectivity and sparsity; the coarse-grained graph generates region nodes through superpixel segmentation and defines edges based on spatial adjacency. The two-layered graph is connected through a mapping function, providing a rich structured foundation for subsequent inference.

[0051] 4) Implement bottom-up topological reasoning for the geometric distribution learner The geometric distribution learner performs multi-scale information propagation on the hierarchical graph: First, an N-layer Graph Attention Network (GAT) is applied to the fine-grained graph to learn the micro-interaction patterns between pathological feature points; then, pooling operations are used to aggregate the enhanced pathological feature point features to the corresponding region nodes, updating the node representations of the coarse-grained graph; next, an M-layer GAT is applied to the coarse-grained graph to model macro-spatial distribution patterns (such as symmetry and radial growth); finally, a compact spatial distribution vector is generated through a global graph readout mechanism. This vector encodes the multi-scale geometric topological properties of the lesions, providing structured evidence for classification decisions.

[0052] 5) Four-stage progressive resolution analysis and heterogeneous feature fusion The framework employs a four-stage sequence, with each stage processing lesion images at progressively decreasing spatial resolution. The first-stage embedding layer uses various convolutional kernels, such as 3×3, 5×5, 9×9, and 27×27, to capture initial lesion features, while subsequent stages use smaller kernels for feature refinement. This multi-stage fusion mechanism ensures the layer-by-layer accumulation of topological knowledge from early global structure to later detailed patterns, building a deep understanding of the lesion structure.

[0053] To address the challenge of existing technologies in handling complex lesion structures and crucial topological relationships that are difficult to model explicitly, this invention proposes a multi-stage decompositional framework: the first stage extracts initial visual features through a saliency map embedding layer; the second to fourth stages progressively construct a hierarchical map of pathological feature points and learn their spatial distribution patterns through cascaded PSG blocks; and the topological vectors generated in each stage are fed back into the visual feature flow through an attention fusion mechanism, achieving structure-guided feature enhancement. This design enables the model not only to identify "what" the pathological features are, but also to understand "where" and "how" they are spatially organized, significantly narrowing the gap with expert diagnostic logic.

[0054] The specific diagnostic process for the entire model architecture is implemented through the following steps: Step 1: Construct a multi-scale saliency block embedding layer to extract initial visual representations To address the challenges of diverse scales and uneven spatial distribution of pathological features in dermoscopy images, this invention designs an adaptive block embedding strategy. For example... Figure 1 As shown, the initial embedding layer of each stage receives the input image or feature map from the previous stage and generates a salient feature map. The first stage employs a multi-kernel strategy, using convolutional kernels of four sizes—3×3, 5×5, 9×9, and 27×27—in parallel to capture local texture, mesoscopic structure, and global lesion morphology, respectively. The output features are then concatenated and fused. This design ensures the model possesses multi-scale perception capabilities early on, providing rich input representations for subsequent pathological feature point mining. Subsequent stages use smaller kernels (such as 3×3 and 5×5) for feature refinement, reducing computational cost while maintaining feature discriminative power. Figure 1 The overall architecture of the DermPrism model proposed in this invention is shown. This model processes skin lesion images through four consecutive stages. Each stage contains a saliency map embedding layer and a series of pathological-spatial graph (PSG) modules. The PSG module consists of a pathological feature point miner (PCP-Miner), a hierarchical graph connector (H-Connector), and a geometric distribution learner, used to learn the spatial distribution of pathological features. The learned spatial vectors are then re-fused into the main feature map through attention fusion. The final features, along with all spatial vectors, are input into a classification head for diagnosis.

[0055] The specific multi-scale feature fusion formula is as follows:

[0056] in, For the input image or feature map, Indicates the kernel size as Convolution operation, This indicates a splicing operation at the channel dimension.

[0057] Step 2: Introduce a pathological feature point mining tool to achieve intelligent identification of sparse key pathological points. After obtaining the salient feature map in step one, this step transforms the dense feature representation into a sparse, semantically rich set of key points using a pathological feature point miner. Specifically, the importance scoring network generates a pathological correlation map using two 1×1 convolutional layers and a GELU / Sigmoid activation function, quantifying the diagnostic value of each spatial location. A Top-K sampling strategy is used to select K=256 highest-scoring points, whose coordinates form a node set. For each node, features are extracted from three modalities: deep semantic features are directly indexed from the corresponding position in the feature map; color features convert RGB values ​​to CIELab space to better perceive color variations in malignant lesions; and texture features extract local neighborhood patterns using a Gabor filter bank. These three types of features are concatenated to form a composite node feature with dimension D_feat. This sparse sampling mechanism allows the model to focus on key structures, reducing computational complexity from O(HW) to O(K), significantly improving efficiency.

[0058] The specific calculation process of the importance scoring network is as follows:

[0059] in For saliency feature maps, and For two different Convolutional layer This represents the GELU activation function. This refers to the Sigmoid function.

[0060] The formula for extracting heterogeneous features of PCP nodes is:

[0061] in For deep semantic features, Represents color transformation functions (RGB) CIE Lab) This represents the texture extraction function of the Gabor filter bank. For nodes The local neighborhood, This indicates feature splicing.

[0062] Step 3: Construct hierarchical graph connectors to organize multi-scale topological relationships Based on the PCP node features and coordinates obtained in step two, this step constructs a hierarchical graph representation. The fine-grained graph construction employs a four-quadrant minimum distance connection algorithm: for each source node, the plane is divided into four quadrants centered on its coordinates. Within each quadrant, the nearest neighbor node is found based on Manhattan distance, establishing bidirectional edges. This strategy ensures that each node is connected in all four directions, avoiding the isolated subgraph problem that may occur with standard k-NN methods and supporting information propagation in long-range structures (such as peripheral stripes). Simultaneously, the coarse-grained graph generates region nodes by superpixel segmentation of image patches, defining edge sets based on spatial adjacency. A mapping function assigns each PCP to a corresponding region node, establishing cross-hierarchical associations. The final output hierarchical graph G_h is a binary tuple (G_fine, G_coarse, M), simultaneously supporting micro and macro topological analysis.

[0063] The formula for constructing the edge set of the minimum distance connection in the four quadrants is: in Relative to node The The set of nodes in each quadrant Distance to Manhattan:

[0064] Step 4: Implement hierarchical topological reasoning and vector distillation for the geometric distribution learner On the hierarchical graph constructed in step three, the geometric distribution learner performs a bottom-up inference process. First, an N=4-layer Graph Attention Network (GAT) is applied to the fine-grained graph to dynamically update the features of pathological feature nodes and learn the contextual interactions between local pathological points. The attention weights of the GAT layers are calculated as follows:

[0065] in For nodes Features The weight matrix is ​​a learnable matrix. For attention weight vectors, This indicates a splicing operation. For nodes The set of neighbors.

[0066] Subsequently, the enhanced PCP features are aggregated to the corresponding region nodes through pooling operations to generate region-enhanced features:

[0067] AGGREGATE can use permutation-invariant functions such as mean, maximum, or summation.

[0068] Next, an M=3-layer GAT is applied to the coarse-grained map to model macroscopic spatial distribution patterns (such as symmetry and radial growth). Finally, a global graph readout mechanism is applied to perform global average pooling on the final region node features and input them into a multilayer perceptron (MLP) to generate a spatial distribution vector D of dimension D_dist. This vector compactly encodes the multi-scale geometric and topological attributes of the lesions, providing a structured basis for classification decisions.

[0069] Step 5: Define the attention fusion mechanism and classification head to implement topology-guided diagnosis. In the aforementioned stages, the spatial distribution vector generated by the geometric distribution learner Injecting attention into the main visual flow through an attention fusion mechanism. Specifically, First, it is expanded to include saliency feature maps. With the same spatial dimension, weights are then calculated using a channel attention mechanism:

[0070] in Indicates a fully connected layer. It is the ReLU activation function. The sigmoid function is used. The fused features are:

[0071] in This represents element-wise multiplication. The process is executed iteratively in four phases, ensuring that deep features are continuously guided by structural understanding.

[0072] After the final stage, the enhanced feature map and the set of all spatial distribution vectors are... The classification head is input together with the lesion type. The classification head consists of a global average pooling layer and a fully connected layer, and outputs the predicted probability of the lesion type. Training uses cross-entropy loss:

[0073] in For real labels, To predict probabilities, This represents the number of categories.

[0074] Step Six: Experimental Verification and Performance Evaluation To verify the effectiveness of this invention, it was evaluated on the ISIC2020 dataset (33,125 images, 8 types of lesions) and the Derm7pt dataset (1,011 cases, including 7-point checklist fine-grained annotations). Figure 2The visualization results of pathological feature points identified by the DermPrism model are presented. Column (a) shows the original skin lesion image, column (b) presents clinical annotations for comparison, and column (c) uses colored dots to mark key areas of interest to the model, specifically reflecting the type and spatial distribution of different feature points. The seven typical dermoscopic features covered in the figure include: PN (papulonodular structure), BWV (blue-white veil), VS (vascular structure), PIG (pigmentation), STR (stripe), DaG (dot-spherical structure), and RS (regressive structure). Visualization analysis shows that the model can effectively simulate the diagnostic thinking of clinicians, accurately distinguish different pathological features, and thus achieve accurate prediction of skin disease types. Figure 2 As shown, qualitative visualization demonstrates a high degree of agreement between the keypoints identified by the PCP miner and clinically labeled high-risk structures (such as blue-white curtains, pigmented areas, and stripes). Table 1 shows the quantitative comparison results, indicating that DermPrism achieved a Balanced Accuracy (BAC) of 89.38% and an AUC of 82.14% on the ISIC2020 dataset, representing improvements of 1.67% and 1.89% respectively compared to the current best-performing comparison method, MambaOut-T. Table 2 shows that on the Derm7pt dataset, DermPrism achieved BAC and AUC of 89.99% and 88.10% respectively, representing improvements of 4.27% and 4.07% compared to the best-performing comparison method, Xception.

[0075]

[0076] Table 1 shows the classification performance of different methods on the ISIC2020 dataset. The evaluation metrics are: accuracy, precision, sensitivity, specificity, F1 score, balanced accuracy, Kappa coefficient, and area under the ROC curve. Table 2 shows the classification performance of different methods on the Derm7pt dataset.

[0077]

[0078] Furthermore, the embodiments of the present invention may also have the following alternatives: (1) Alternative implementation of PCP miner: The importance scoring network can be replaced by a lightweight CNN (such as MobileNet blocks) or a Transformer-based scorer instead of a 1×1 convolutional structure; Top-K sampling can be replaced by differentiable sampling (such as Gumbel-Softmax) to support end-to-end gradient propagation, but may introduce randomness.

[0079] (2) Alternative graph construction strategies: Four-quadrant minimum distance connection can be replaced by k-nearest neighbor (k-NN) based on Euclidean distance or radius graph, but may produce isolated nodes; superpixel segmentation can be replaced by region generation method based on deep clustering, but the computational cost is greater.

[0080] (3) Replacement of geometric distribution learners: The GAT layer can be replaced with GraphSAGE or GCN to reduce the computational complexity of attention, but may lose the ability to dynamically adjust weights; the global readout mechanism can use more complex virtual nodes or attention pooling to replace simple global pooling + MLP.

[0081] (4) Replacement of multi-stage architecture: The four-stage design can be reduced to three stages or extended to five stages to balance accuracy and efficiency; the fusion between stages can be replaced by simple feature splicing instead of attention fusion, but the topology guidance role is weakened.

[0082] (5) Modal feature substitution: CIELab color space can be replaced with HSV or YCbCr; Gabor filter bank can be replaced with LBP or HOG texture descriptor, but may change feature discrimination characteristics.

[0083] In summary, this invention achieves explicit decomposition and learning of pathological feature topology through pathological spatial tiles, systematically solving the shortcomings of existing technologies in structural modeling, computational allocation, and graph representation, and significantly improving diagnostic accuracy, robustness, and clinical interpretability.

[0084] It should be emphasized that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A pathological spatial map modeling system for intelligent diagnosis of skin lesions, characterized in that, include: A pathological feature point miner is used to generate pathological correlation maps through a lightweight scoring network and selects a fixed number of K most diagnostically valuable pathological feature points using a deterministic Top-K sampling strategy. Layered graph connectors are used to construct composite graph structures; Geometric distribution learners are used to perform multi-scale information propagation on hierarchical graphs. The pathological spatial map modeling system includes four parallel deployments of visual representation learning and topology modeling. The parallel deployments of visual representation learning and topology modeling include a saliency map embedding layer and multiple pathological spatial map tiles. The pathological spatial map tiles are sequentially composed of the pathological feature point miner, the hierarchical map connector, and the geometric distribution learner. The pathological spatial tiles transform lesion images into clinically inspired structured representations; The parallel deployment of visual representation learning and topology modeling uses an attention fusion mechanism to iteratively inject the learned spatial distribution vectors into the main visual stream, so that the depth visual features are continuously guided by topological understanding, thereby achieving synergistic enhancement of the two types of information.

2. The pathological spatial map modeling system for intelligent diagnosis of skin lesions according to claim 1, characterized in that, The pathological feature point mining tool concatenates heterogeneous features extracted from deep semantic feature maps, CIELab color space transformation, and Gabor filter banks for each pathological feature point to construct rich composite feature vectors; it concentrates computational resources on clinically significant structures, reducing the complexity of graph construction while retaining key diagnostic information; When the layered graph connector constructs a composite graph structure, the fine-grained graph uses PCP as nodes and adopts a four-quadrant minimum distance connection algorithm to establish edges—for each source node, the nearest neighbor node with Manhattan distance is found in the coordinate quadrants; the coarse-grained graph generates region nodes through superpixel segmentation and defines edges based on spatial adjacency. The geometric distribution learner first applies an N-layer graph attention network to learn the micro-interaction patterns between pathological feature points on the fine-grained graph; then, it aggregates the enhanced pathological feature point features to the corresponding region nodes through pooling operations, updating the node representation of the coarse-grained graph; next, it applies an M-layer GAT to model the macro-spatial distribution pattern on the coarse-grained graph; finally, it generates a compact spatial distribution vector through a global graph readout mechanism.

3. The pathological spatial map modeling system for intelligent diagnosis of skin lesions according to claim 2, characterized in that, The pathological spatial map modeling system adopts a multi-stage decomposition framework: the first stage extracts initial visual features through a saliency map embedding layer; In the second to fourth stages, a hierarchical map of pathological feature points is gradually constructed through cascaded pathological spatial tiles, and their spatial distribution patterns are learned. The topological vectors generated in each stage feed back into the visual feature flow through an attention fusion mechanism, thereby achieving feature enhancement guided by structure.

4. A pathological spatial map modeling method for intelligent diagnosis of skin lesions, characterized in that, The modeling method is used in the modeling system according to any one of claims 1 to 3, and the modeling method includes the following steps: Step 1: Construct a multi-scale saliency block embedding layer to extract the initial visual representation; Step 2: Introduce a pathological feature point mining tool to achieve intelligent identification of sparse key pathological points; Step 3: Construct hierarchical graph connectors to organize multi-scale topological relationships; Step 4: Implement hierarchical topological reasoning and vector distillation for the geometric distribution learner; Step 5: Define the attention fusion mechanism and classification head to implement topology-guided diagnosis.

5. The pathological spatial map modeling method for intelligent diagnosis of skin lesions according to claim 4, characterized in that, Step one specifically includes: The initial embedding layer of each stage receives the input image or the feature map from the previous stage and generates a salient feature map. The first stage uses a multi-convolution kernel strategy to capture local texture, mesoscopic structure and global lesion morphology respectively, and the output features are fused by concatenation. Subsequent stages use smaller kernels to refine the features. The multi-scale feature fusion formula is as follows: in, For the input image or feature map, Indicates the kernel size as Convolution operation, This indicates a splicing operation at the channel dimension.

6. The pathological spatial map modeling method for intelligent diagnosis of skin lesions according to claim 4, characterized in that, Step two specifically includes: The dense feature representation is transformed into a sparse, semantically rich set of key points through a pathological feature point mining tool; the importance scoring network generates a pathological correlation map through convolutional layers and GELU / Sigmoid activation function to quantify the diagnostic value of each spatial location; the highest score point is selected by a Top-K sampling strategy, and its coordinates constitute a set of nodes. For each node, features are extracted from three modalities: deep semantic features are directly indexed from the corresponding positions in the feature map, color features are converted from RGB values ​​to CIELab space, and texture features are extracted from local neighborhood patterns through Gabor filter banks; the three types of features are concatenated to form a composite node feature with dimension D_feat; The specific calculation process of the importance scoring network is as follows: in, For saliency feature maps, and For two different Convolutional layer This represents the GELU activation function. For the Sigmoid function; The formula for extracting heterogeneous features of PCP nodes is: in For deep semantic features, This represents a color transformation function. This represents the texture extraction function of the Gabor filter bank. For nodes The local neighborhood, Indicates feature splicing, This represents the input image for the k-th image patch.

7. The pathological spatial map modeling method for intelligent diagnosis of skin lesions according to claim 4, characterized in that, Step three specifically includes: Fine-grained graph construction employs a four-quadrant minimum distance connection algorithm: for each source node, the plane is divided into four quadrants centered on its coordinates, and the nearest neighbor node is found in each quadrant based on the Manhattan distance, and bidirectional edges are established; Meanwhile, the coarse-grained graph generates region nodes by superpixel segmentation of image patches and defines edge sets based on spatial adjacency; the mapping function assigns each PCP to the corresponding region node to establish cross-level associations; the final output hierarchical graph G_h is a pair of tuples, which simultaneously supports micro and macro topological analysis. The formula for constructing the edge set of the minimum distance connection in the four quadrants is: in Relative to node The The set of nodes in each quadrant For Manhattan distance, The value of the variable that makes the function reach its minimum value. Represents the k-th feature node or vertex. Represents the set of all selected fine-grained feature nodes. Represents the x-coordinate of the i-th feature node in the image. Represents the x-coordinate of the k-th feature node in the image. Represents the ordinate of the i-th feature node in the image, This represents the ordinate of the k-th feature node in the image.

8. The pathological spatial map modeling method for intelligent diagnosis of skin lesions according to claim 4, characterized in that, Step four specifically includes: The geometrically distributed learner performs a bottom-up inference process; firstly, an N=4-layer graph attention network is applied to the fine-grained graph to dynamically update the features of pathological nodes and learn the contextual interactions between local pathological points; the attention weights of the GAT layer are calculated as follows: in, For nodes Features The weight matrix is ​​a learnable matrix. For attention weight vectors, This indicates a splicing operation. For nodes The neighborhood group, For activation function, The feature vector of node j The feature vector of node k The transpose operation represents a matrix or vector. Represents the natural exponential function; Subsequently, the enhanced PCP features are aggregated to the corresponding region nodes through pooling operations, generating the region-enhanced features as follows: in, This indicates the aggregation of a set of node features. This function represents mapping node vi to its region. The identifier of the j-th region or supernode This represents the updated feature vector of the i-th node at the current stage. Indicates the index of the node; AGGREGATE can use permutation-invariant functions such as mean, maximum, or summation; Next, an M=3 layer GAT is applied to the coarse-grained map to model the macroscopic spatial distribution pattern; finally, a global graph readout mechanism is applied to perform global average pooling on the final region node features and input them into a multilayer perceptron to generate a spatial distribution vector D with dimension D_dist.

9. A pathological spatial map modeling method for intelligent diagnosis of skin lesions according to claim 4, characterized in that, Step five specifically includes: First, it is expanded to include saliency feature maps. With the same spatial dimension, weights are then calculated using a channel attention mechanism: in, Indicates a fully connected layer. It is the ReLU activation function. For the Sigmoid function, This represents the original feature vector or feature map of the i-th branch or stage; the fused features are: in, This represents element-wise multiplication. This represents the deep feature map extracted by the backbone network; the process is executed iteratively in four stages, ensuring that the deep features are continuously guided by structural understanding. After the final stage, the enhanced feature map and the set of all spatial distribution vectors are... The classification head is input together; the classification head consists of a global average pooling layer and a fully connected layer, and outputs the predicted probability of lesion type. Training uses cross-entropy loss: in, For real labels, To predict probabilities, This represents the number of categories.

Images