Microsoft Research and the University of Alicante announce PadChest-GR, a groundbreaking new benchmark for Grounded Radiology Reporting. This collaboration, involving University Hospital Sant Joan d’Alacant and MedBravo, introduces the world’s first multimodal, bilingual sentence-level radiology report dataset. PadChest-GR redefines how AI and radiologists interpret radiological images, setting a new standard for AI in healthcare.
Traditional radiology reports often provide unstructured narratives. In contrast, grounded reporting demands individual description and localization for each finding. This approach mitigates AI fabrication risks and enhances clinical interpretability. PadChest-GR addresses this need with 4,555 chest X-ray studies. These studies feature Spanish and English sentence-level descriptions, plus precise spatial (bounding box) annotations for both positive and negative findings. It is the first public benchmark enabling evaluation of fully grounded radiology reports in chest X-rays.
PadChest-GR: Driving Interpretable AI
PadChest-GR plays a critical role in powering Microsoft's state-of-the-art multimodal report generation model, MAIRA-2. Leveraging PadChest-GR's detailed annotations, MAIRA-2 exemplifies a commitment to building more interpretable and clinically useful AI systems. This initiative highlights the transformative power of interdisciplinary collaboration. The University of Alicante provided deep clinical expertise, enriching the dataset's relevance and utility.
The annotation process for PadChest-GR integrated advanced large language models with comprehensive manual review. Microsoft Azure OpenAI Service, utilizing GPT-4, extracted and translated sentences from raw reports. It also linked these to existing expert labels. Subsequently, radiologists from Hospital San Juan de Alicante performed meticulous quality checks on the Centaur Labs platform. They annotated each positive finding with bounding boxes, ensuring accuracy and consistency.
PadChest-GR not only establishes a new benchmark but also serves as a foundational dataset for MAIRA-2. This dataset is openly available, encouraging the broader research community to build upon it. Fostering open collaboration and sharing resources accelerates progress in medical imaging AI. Ultimately, this improves patient care. Researchers and industry experts can explore PadChest-GR and MAIRA-2, contributing to the advancement of grounded radiology reporting.