Search papers, labs, and topics across Lattice.
This study introduces a dual-level relational framework for skin lesion diagnosis that integrates implicit representation learning with explicit structural modeling through graph-based methods. By employing a convolutional masked autoencoder for implicit inter-patch relationship learning and organizing embeddings into various graph topologies, the authors demonstrate significant performance improvements on the ISIC-2018 and ISIC-2019 benchmarks. The best results were achieved with a grid-structured Graph Attention Network, which improved accuracy from 76.17% to 79.27% on ISIC-2018, showcasing the effectiveness of combining these relational biases.
Bridging implicit and explicit relational biases boosts skin lesion diagnosis accuracy by nearly 3% using graph-based methods.
Relational inductive biases are essential for capturing structural dependencies among data. This study investigates a dual-level relational framework for image classification, bridging the gap between implicit representation learning and explicit structural modelling. We begin by establishing a baseline using an EfficientNetB3 architecture. To move beyond standard convolutional biases, we adopt a patch-based strategy, employing a convolutional masked autoencoder to learn implicit inter-patch relationships through self-supervised reconstruction. We then extend this approach by incorporating explicit relational modelling, organizing the learned embeddings into various graph topologies, including grid-based, random, and k-nearest neighbour structures. Experimental results on the ISIC-2018 and ISIC-2019 skin lesion diagnosis benchmarks show that combining implicit inter-patch modelling with explicit graph-based message passing yields the best performance. On the ISIC-2018 test set, the baseline model achieves a balanced accuracy of 76.17%, which improves to 77.12% with implicit patch-based relational modelling. The fully integrated grid-structured Graph Attention Network further increases performance to 79.27%. Similarly, on ISIC-2019, the implicit approach reaches 59.84% balanced accuracy, while the combination of implicit and explicit modelling yields 60.67%.
