RadEdit: A deep learning biomedical vision models to discover failure cases

About Model

RadEdit is a deep learning-based text-to-image diffusion model developed by Microsoft Health Futures for stress-testing biomedical vision models. It enables the editing of chest X-ray images by adding or removing abnormalities based on text descriptions, allowing researchers to evaluate AI models for disease classification, anatomy segmentation, and robustness in various clinical scenarios. Trained on 487,680 chest X-ray images from datasets such as MIMIC-CXR, NIH-CXR, and CheXpert, RadEdit can: 1. Modify X-rays based on text prompts (e.g., adding "pleural effusion" or removing "pneumothorax"). 2. Generate synthetic X-rays conditioned on a radiology report or list of abnormalities. 3. Assess AI model robustness by testing performance across different medical conditions. The model is powered by latent diffusion, integrating components such as BioViL-T (text encoder) and SDXL-VAE (autoencoder) for high-quality image editing. RadEdit is intended for research purposes only and is not suitable for clinical use, as it may introduce biases and errors. It serves as a valuable tool for AI researchers exploring failure cases, dataset shifts, and model generalizability in medical imaging.