RAD-DINO-MAIRA-2 - Vision Transformer for Medical Imaging

About Model

RAD-DINO-MAIRA-2 is a vision transformer model designed for encoding chest X-ray images using self-supervised learning based on the DINOv2 framework. It is an enhanced version of RAD-DINO, trained on a larger dataset to improve its performance in medical imaging tasks. Developed by Microsoft Health Futures, this model serves as the vision backbone for MAIRA-2, facilitating radiology report generation by extracting meaningful image features. The model can be used for various downstream applications, including: 1. Image classification with a classifier trained on extracted embeddings. 2. Image segmentation using a decoder trained on patch tokens. 3. Medical image retrieval via nearest-neighbor search. 4. Clustering using learned image embeddings. 5. Radiology report generation by integrating with a language model. Trained on 1.4 million chest X-ray images from datasets such as MIMIC-CXR, NIH-CXR, PadChest, and CheXpert, the model provides a robust representation of medical images. However, RAD-DINO-MAIRA-2 is for research purposes only and is not intended for clinical use. It enables advancements in AI-powered medical imaging but requires careful validation before deployment in healthcare applications.