Reproduce “ProtoMIL: Multiple Instance Learning with Prototypical Parts for Whole-Slide Image Classification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dawid Rymarczyk, Adam Pardyl, Jarosław Kraus, Aneta Kaczyńska, Marek Skomorowski, Bartosz Zieliński
ProtoMIL introduces a Multiple Instance Learning framework that provides fine-grained, case-based explanations for Whole-Slide Image classification while achieving competitive performance, including an AUC of 0.935 on the Camelyon16 dataset.
Key insights
- Combining attention-based MIL with a prototype layer effectively generates fine-grained, case-based explanations by identifying crucial patches and comparing them to visually interpretable prototypes.
- Leveraging attention weights within regularization losses successfully guides the learning of discriminative prototypes, particularly for sparse positive instances, without compromising overall performance.
- The approach provides both local (explaining a single WSI prediction) and global (showing all learned prototypes) interpretability, thereby bridging the gap between AI predictions and clinical understanding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.