Deep learning models for computational pathology often struggle to explain whole-slide image (WSI) predictions in terms that pathologists can use. Existing methods tend to highlight important regions without identifying the morphological patterns behind them. ProtoGTX addresses this by building a graph-based framework that adds a flexible morphological prototype vocabulary to spatial context.
The method separates the reference prototypes—which define morphological concepts—from the predictive model. The reference vocabulary can be generated from training data via unsupervised clustering or supplied externally, and it stays fixed during classifier training. Patch features are first contextualized over a spatial graph, then learnable prototypes are optimized through cross-attention, and the slide-level activation profile is aligned with the morphological composition from the fixed reference bank.
On the CPTAC and BRACS datasets, ProtoGTX outperformed the evaluated MIL, prototype-based, and graph-based baselines, with especially strong gains on BRACS. Qualitative analyses show that spatial contextualization turns fragmented patch-level prototype assignments into more coherent tissue regions, and that prototype maps co-localize with pathologist annotations and regions relevant to slide-level prediction. This suggests ProtoGTX can provide complementary semantic and spatial explanations while keeping the prototype vocabulary independent of the predictive model.