MedSAE: Dissecting MedCLIP Representations with Sparse Autoencoders
Fuente:
arXiv
Guardado en:
| Autores principales: | , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2025
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866916037475172352 |
|---|---|
| author | Renzulli, Riccardo Lepoutre, Colas Cassano, Enrico Grangetto, Marco |
| author_facet | Renzulli, Riccardo Lepoutre, Colas Cassano, Enrico Grangetto, Marco |
| contents | Artificial intelligence in healthcare requires models that are accurate and interpretable. We advance mechanistic interpretability in medical vision by applying Medical Sparse Autoencoders (MedSAEs) to the latent space of MedCLIP, a vision-language model trained on chest radiographs and reports. To quantify interpretability, we propose an evaluation framework that combines correlation metrics, entropy analyses, and automated neuron naming via the MedGemma foundation model. Experiments on the CheXpert dataset show that MedSAE neurons achieve higher monosemanticity and interpretability than raw MedCLIP features. Our findings bridge high-performing medical AI and transparency, offering a scalable step toward clinically reliable representations. The source code supporting the findings of this study is available at https://github.com/EIDOSLAB/MedSAE. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2510_26411 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | MedSAE: Dissecting MedCLIP Representations with Sparse Autoencoders Renzulli, Riccardo Lepoutre, Colas Cassano, Enrico Grangetto, Marco Artificial Intelligence Artificial intelligence in healthcare requires models that are accurate and interpretable. We advance mechanistic interpretability in medical vision by applying Medical Sparse Autoencoders (MedSAEs) to the latent space of MedCLIP, a vision-language model trained on chest radiographs and reports. To quantify interpretability, we propose an evaluation framework that combines correlation metrics, entropy analyses, and automated neuron naming via the MedGemma foundation model. Experiments on the CheXpert dataset show that MedSAE neurons achieve higher monosemanticity and interpretability than raw MedCLIP features. Our findings bridge high-performing medical AI and transparency, offering a scalable step toward clinically reliable representations. The source code supporting the findings of this study is available at https://github.com/EIDOSLAB/MedSAE. |
| title | MedSAE: Dissecting MedCLIP Representations with Sparse Autoencoders |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2510.26411 |