Topology-Guided Graph-Visual Fusion for Interpretable Skin Lesion Diagnosis

Accurate skin lesion diagnosis requires interpreting the complex spatial distributions and topological relationships of pathological structures. However, prevalent deep learning paradigms process images as regular grids or flattened sequences, severely limiting their capacity to explicitly model these non-Euclidean clinical priors. To bridge this semantic gap, we propose DermPrism, a topology-guided graph-visual fusion framework with distinct visual and topological pathways that are coupled within one predictive network. Central to our approach is a novel architecture that dynamically mines sparse, diagnostically salient keypoints, structuring them into a hierarchical graph. This enables explicit geometric reasoning over both microscopic feature interactions and macroscopic regional layouts. Furthermore, a dedicated modulation mechanism seamlessly injects these learned topological priors back into the dense visual representations. Comprehensive evaluations on multiple benchmarks show that DermPrism achieves state-of-the-art performance and improved cross-dataset transfer. Its visualizations expose hierarchical topology and attention flow that are qualitatively consistent with annotated dermoscopic structures, without implying equivalence to dermatologist reasoning.