Study reveals encoder choice drives graph homophily in breast ultrasound AI

A new study from researchers including Sabahattin Mert Daloglu, Ceren Coskun, Harvey Castro, Soner Hacihaliloglu, and Ilker Hacihaliloglu systematically analyzed how the choice of image encoder affects graph-based breast ultrasound classification. The team tested five encoders spanning convolutional and transformer architectures, using their embeddings to construct cosine similarity k-nearest-neighbor graphs. These graphs were classified with a single-layer graph convolutional network (GCN) and a linear classification head. Across three patientwise cross-validation folds, they found that higher-capacity encoders consistently improved graph homophily—the tendency for nodes of the same class to be connected—and downstream metrics including accuracy, AUC, sensitivity, specificity, and F1-score.

Notably, the study revealed a strong linear correlation between test-set graph homophily and classification accuracy, with higher-capacity encoders consistently occupying the high-homophily, high-accuracy region. This suggests that encoder-driven improvements in graph structure are a key mechanism behind the performance gains, not just better feature extraction alone. The work, submitted to the MICCAI 2026 ASMUS Workshop, positions encoder selection as a critical design choice for graph-based medical imaging systems and identifies graph homophily as a proxy metric for representation quality. These insights could help streamline the development of more accurate and reliable automated breast ultrasound screening tools.