Embedding Anisotropy in Medical Image Foundation Models: Code and Data

Code and precomputed embeddings to reproduce all experiments in "Embedding Anisotropy in Medical Image Foundation Models: Impact of Post-Hoc Whitening on Downstream Classification." This repository contains: A self-contained Python pipeline (reproduce_experiments.py) that extracts embeddings from 4 foundation models (ResNet50, DINOv2-ViTS14, BiomedCLIP, PLIP) across 6 MedMNIST benchmarks, computes anisotropy metrics, runs classification experiments under 4 whitening conditions, and performs statistical tests. A figure generation script (generate_figures.py) producing all publication figures. Precomputed embedding matrices (24 .npz files) enabling full reproduction without GPU access. All experimental results as CSV files. See README.md for full instructions.