Gdańsk Tech's graph AI lifts skin lesion diagnosis to 79.3% accuracy

A team from Gdańsk University of Technology and Medical University of Gdańsk has introduced a novel dual-level relational framework for image classification, specifically targeting skin lesion diagnosis. Their paper, accepted at CIBB 2026, tackles the challenge of integrating implicit representation learning with explicit structural modeling. The researchers start with an EfficientNetB3 baseline, then move beyond standard convolutional biases using a patch-based convolutional masked autoencoder that learns inter-patch relationships through self-supervised reconstruction. This implicit modeling alone improved balanced accuracy on ISIC-2018 from 76.17% to 77.12%.

To push further, the team added explicit relational modeling by organizing learned embeddings into various graph topologies—grid, random, and k-nearest neighbor—and applying graph attention networks for message passing. The fully integrated grid-structured Graph Attention Network achieved the best results: 79.27% balanced accuracy on ISIC-2018 and 60.67% on ISIC-2019. This research demonstrates that combining implicit patch relationships with explicit graph structures can meaningfully enhance medical image classification, potentially leading to more reliable automated dermatology tools.