Hybrid Graph Attention Framework for Dermoscopic Classification via Topology-Aware Latent Embeddings
Year: 2026
Type: article
Abstract: Background: Automated stratification of pigmented skin lesions remains a formidable challenge in computational pathology due to high inter-class visual similarity and severe class imbalance. While Vision Transformers (ViTs) have established new benchmarks by capturing global context, they typically process image patches as independent tokens, often failing to explicitly model the topological and spatial relationships critical for differentiating complex lesion morphologies.Methods: We propose a hybrid framework that synergizes self-supervised feature extraction with geometric deep learning to model the latent structural topology of dermoscopic images. The architecture leverages a pre-trained foundation model as a backbone, integrated with low-rank adaptation layers for domain-specific refinement. Unlike conventional grid-based methods, our approach dynamically constructs a k-Nearest Neighbor (k-NN) graph from high-dimensional patch embeddings. This latent graph is then processed through a Graph Attention Network (GAT), enabling the model to explicitly reason over spatial relationships and multi-scale structural dependencies.Results: Evaluated on the HAM10000 benchmark, Under identical experimental conditions, our method achieves an accuracy of 97.80% and an AUC of 99.65%, outperforming both CNN- and Transformer-based baselines. Notably, this performance is achieved with high computational efficiency, utilizing only 3.19% trainable parameters (10M) and demonstrating rapid convergence within the first epoch.Conclusion: This study demonstrates that explicit modeling of spatial interactions via GNNs significantly enhances diagnostic precision in dermoscopy without the need for massive retraining. Although our experiments focus on HAM10000, the design of the framework is expected to generalize to broader dermoscopic datasets due to its reliance on SSL features and topology-aware reasoning. The proposed architecture offers a scalable, interpretable, and clinically viable solution for computer-aided diagnosis systems.
DOI: 10.7240/jeps.1862463
Author: Tolga Saim Başçetin
Institution: Yıldız Technical University
Language: Turkish
Subfield: Oncology
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W7214559819