Bag-of-Visual-Words for Spatial Mapping of Lung Adenocarcinoma Growth Patterns

The paper presents a weakly supervised Bag-of-Visual-Words (BoVW) pipeline for mapping lung adenocarcinoma (LUAD) growth patterns across whole-slide images. It learns a visual vocabulary from frozen foundation-model embeddings extracted from a small set of annotated regions of interest, builds class prototypes as mean BoVW histograms, and classifies sliding-window regions using Jensen–Shannon divergence. On 87 CPTAC-LUAD patients, the best tumour/healthy configuration reached 0.974 balanced accuracy with H-Optimus-1, close to the 0.987 supervised SVM baseline using mean-pooled WSI embeddings. For binary histologic grade, BoVW outperformed that baseline with every tested encoder.