Hybrid Graph-Based MIL Improves Skin Lesion Diagnosis Performance
August 7, 2026
A dual-level framework combines EfficientNetB3 with a convolutional masked autoencoder for implicit patch relationships and graph-based message passing for explicit structural modeling. Testing on ISIC-2018 and ISIC-2019 benchmarks shows that integrating these relational biases outperforms standard convolutional approaches.
HOW THIS AFFECTS YOU
●
researcherYou can leverage this dual-level architecture to better capture structural dependencies in medical imaging.
●
healthThis method improves the accuracy of automated skin lesion diagnosis using graph-based modeling.