Search papers, labs, and topics across Lattice.
This paper introduces MRI2Rep, an autoregressive framework designed for the automated generation of structured reports from 3D liver MRI scans, addressing the challenges posed by volumetric complexity and limited paired data. Utilizing a Report-to-Label Canonicalization (RLC) module, the system converts free-text reports into structured diagnostic sequences without requiring lesion-level annotations, achieving significant improvements in case-level sensitivity and liver-level accuracy compared to existing medical vision-language models. In a blinded reader study, radiologists rated a substantial proportion of the AI-generated reports as clinically acceptable, highlighting the potential of MRI2Rep to streamline the reporting process in radiology.
AI-generated liver MRI reports can achieve 76% case-level sensitivity and are rated clinically acceptable by radiologists, challenging the status quo of manual reporting.
Manual reporting of 3D MRI studies is time-consuming, yet end-to-end structured report generation for 3D liver MRI remains underexplored due to volumetric complexity and scarce paired data. We propose MRI2Rep, an autoregressive framework for liver MRI report generation. From 3,929 real-world MRI-report pairs acquired over a 10-year single-institution cohort, a Report-to-Label Canonicalization (RLC) module converts free-text reports into structured, closed-vocabulary diagnostic sequences without lesion-level annotations. On a held-out test set, MRI2Rep achieves 76.0% case-level sensitivity, 29.4% lesion-level F1, compared with no more than 8.3% for adapted medical vision-language baselines, and 82.4% liver-level accuracy. In a blinded reader study, two radiologists rated 75% and 70% of AI-generated reports as clinically acceptable, compared with 95% and 100% for original reports. Our automated LLM-based judge, LLM-Eval, rated 61.8% of AI-generated reports as acceptable, applying a stricter standard and supporting its use as a conservative proxy. To our knowledge, this is the first end-to-end LI-RADS-structured reporting system for 3D liver MRI.
