MedSAE Improves Interpretability of MedCLIP for Chest Radiographs

A team of researchers has introduced Medical Sparse Autoencoders (MedSAEs) to improve the interpretability of MedCLIP, a vision-language model specifically trained on chest radiographs and their corresponding reports. This research, which appears on arXiv, suggests an evaluation framework that integrates correlation metrics, entropy analyses, and automated neuron naming through the MedGemma foundation model. Tests conducted on the CheXpert dataset reveal that neurons from MedSAE provide greater monosemanticity and interpretability compared to the original MedCLIP features. This advancement connects high-performing medical AI with transparency, representing a scalable progression toward clinically trustworthy representations. The source code is accessible to the public.