Back

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

Cen, L.; Porembka, J. H.; Hayes, J. C.; Merchant, K.; Igboagi, U.; Srivastava, S.; Mootz, A. R.; Topper, V.; Hayes, S.; Yadu, N.; Arjmandi, F. K.; Schopp, J. G.; Dogan, B. E.; Hu, T.

2026-08-05 radiology and imaging
10.64898/2026.08.03.26359608 medRxiv
Show abstract

Existing radiological artificial intelligence (AI) systems are difficult to modify, validate, and adapt to new clinical applications. We present a large language model (LLM)-driven agentic framework capable of reconstructing, optimizing, and customizing deep-learning (DL) systems for radiological image analysis using a single consumer-grade PC. The agent reconstructed the missing pre-training model and corrected a clinical reasoning flaw in an example mammography DL workflow. The improved model performance surpassed all 1,687 submitted models in the Radiological Society of North America Breast Cancer AI Challenge. Across international datasets (n>13,000) from US and China, the model demonstrated robust generalizability (AUC: 0.9). In a reader study (n>1,200), the model outperformed radiologists by an absolute AUC margin of 24% on extended follow-up. Our findings demonstrate that LLM-driven agents enable radiologist-guided customization of radiological AI systems on a consumer-grade PC while reducing the technical expertise required for implementation. This work paves the way for accessible radiological AI.

Matching journals

The top 8 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.