Closing the BRACS performance gap through pretraining or superpixel-size optimization

Determine whether pretraining or a better-selected superpixel size can close TAN's performance gap on the BRACS breast-lesion subtyping dataset relative to the stronger foundation-model encoders.

Background

TAN achieves performance close to a from-scratch ResNet18 on BRACS but trails the UNI and CTransPath histopathology foundation models. The study also notes that superpixel size was matched in pixels rather than tissue area, was not tuned, and may contribute to dataset-dependent performance differences. The authors therefore leave unresolved whether external pretraining or selecting a more suitable superpixel size would eliminate the observed BRACS gap.