Uncertainty-Aware Structured Data Extraction from Full CMR Reports via Distilled LLMs
Researchers have developed CMR-EXTR, a new framework designed to convert free-text cardiac magnetic resonance (CMR) reports into structured data. This system not only extracts information but also assigns confidence scores for each data field, aiding in quality control. The framework utilizes a teacher-student distillation pipeline for offline inference and integrates three principles for uncertainty estimation: distribution plausibility, sampling stability, and cross-field consistency. Experiments indicate CMR-EXTR achieves 99.65% variable-level accuracy, marking it as the first CMR-specific extraction system with built-in confidence estimation. AI
IMPACT Enables more efficient and reliable cohort assembly and clinical decision support through automated data extraction from medical reports.