Towards Accessible Radiological Image Analysis via Local Agentic Framework: Validation in Mammography

A research team developed a new system based on a large language model (LLM) that enables physicians to modify and customize artificial intelligence for radiological image analysis using a standard consumer-grade PC. The system was tested on mammography and improved an existing deep learning model by correcting clinical reasoning errors. The improved model outperformed all 1,687 competing models in the Radiological Society of North America's AI Breast Cancer Challenge. When tested on international datasets with over 13,000 cases from the US and China, the model achieved robust generalizability with an AUC of 0.9. In a study with over 1,200 participants, the model outperformed radiologists by 24 percent in AUC measurement on extended follow-up. This work demonstrates that LLM-driven agents enable radiologists to customize radiological AI systems on a standard consumer-grade PC without requiring deep technical expertise.