Official implementation of the paper PIKACHU: Prototypical In-context Knowledge Adaptation for Clinical Heterogeneous Usage

Accepted to MIDL 2026 (poster)

Folder structure

This README reflects the current files in this repository.

pikachu/
├── .git/
├── README.md
├── config.py
├── dataset.py
├── utils.py
├── icl_module.py
├── run_experiments.py
├── baseline_dinov2.py
├── baseline_medclip.py
├── baseline_siglip.py
├── baseline_vit.py
├── icl_dinov2.py
├── icl_medclip.py
├── icl_siglip.py
├── icl_vit.py
└── data/   (currently empty)

File descriptions

• config.py: dataclass-based configuration for data, model, ICL, and output directories.
• dataset.py: dataset classes and dataloader creation for folder-based, CSV-based, and hybrid modes.
• utils.py: metric computation, confusion matrix/ROC plotting, result saving, and baseline-vs-ICL comparison plots.
• icl_module.py: reusable ICL building blocks (cross-attention ICL, prototype ICL, adaptive fusion, weighted kNN).

Baseline scripts

• baseline_dinov2.py: DINOv2 linear-probe baseline.
• baseline_medclip.py: PubMedCLIP/BiomedCLIP zero-shot baseline.
• baseline_siglip.py: SigLIP2 zero-shot baseline.
• baseline_vit.py: ViT baseline.

ICL scripts

• icl_dinov2.py: DINOv2 ProtoNet-style ICL.
• icl_medclip.py: PubMedCLIP ProtoNet-style ICL.
• icl_siglip.py: SigLIP2 ProtoNet-style ICL.
• icl_vit.py: ViT ProtoNet-style ICL.

Known limitation in run_experiments.py

run_experiments.py matches some renamed files (icl_dinov2.py, icl_vit.py), but it still imports CLIP modules that are not present in this repository:

• missing: baseline_clip.py, icl_clip.py
• available related files: baseline_medclip.py, icl_medclip.py

As a result, DINOv2/ViT paths can run, but CLIP steps fail unless imports are updated or matching files are added.

Data expectations

The data/ directory is currently empty. Before running experiments, provide datasets and update paths in config.py (for example: data_root, train_dir, test_dir, csv_file, image_dir).

Typical usage

1. Edit dataset and experiment settings in config.py.
2. Run one script at a time, e.g.:
   • python baseline_dinov2.py
   • python baseline_medclip.py
   • python icl_dinov2.py
3. Check outputs in results/ and checkpoints/ (created automatically).