MAIRA-2 - Multimodal AI for Radiology Report Generation

About Model

MAIRA-2 is a multimodal Transformer model developed by Microsoft Research Health Futures for automated radiology report generation. It processes chest X-ray images and generates structured reports, with or without grounding, by linking findings to specific image regions. The model is built using the RAD-DINO-MAIRA-2 image encoder and the Vicuna-7b-V1.5 language model, enabling advanced medical text generation. MAIRA-2 is intended for research purposes only and is not suitable for clinical practice due to potential biases and limitations in generalizability. Trained on datasets such as MIMIC-CXR, PadChest, and USMix, it aims to facilitate comparative studies in AI-driven radiology analysis. The model supports findings generation, phrase grounding, and structured medical text generation, providing valuable insights into automated medical imaging interpretation while promoting fairness and responsible AI use in healthcare.