1/ Exciting new work from our team at Harvard! We developed a variant of direct preference optimization (DPO) to fix one of the biggest problems with radiology report generation models - hallucinating references to non-existent prior exams. 🩺📄 -> link to paper at end of thread

When should we trust a radiology report generation model? When the metrics say to. But when should we trust the metrics? New paper ✨ ReXamine-Global, a framework for stress-testing metrics used to automatically score AI-generated radiology reports. Oishi Banerjee