cmrextr-1b

CMR-EXTR: Structured Extraction from Cardiac MRI Reports

CMR-EXTR is a lightweight framework for converting free-text cardiac magnetic resonance (CMR) reports into structured, auditable data with per-field confidence estimation. It was introduced in the paper Uncertainty-Aware Structured Data Extraction from Full CMR Reports via Distilled LLMs.

Overview

The model is designed to support cohort assembly, longitudinal data curation, and clinical decision support in real-world clinical workflows. It performs structured information extraction from reports and assigns confidence scores to each extracted field, enabling efficient human review and quality control.

---

Key Features

• Structured Extraction: Converts free-text CMR reports into predefined structured fields
• Per-field Confidence: Provides uncertainty estimates for each extracted variable
• Offline Inference: Fully deployable without external API dependencies
• Efficient Design: Lightweight student model distilled from a larger teacher model

---

Code

The official implementation is available on GitHub:
CMR-EXTR

---

Method Summary

CMR-EXTR is built on a teacher–student distillation framework:

• A large teacher model generates high-quality structured outputs
• A compact student model (based on Llama-3.2-1B) is trained to replicate these outputs efficiently
• The student model supports fast and fully offline inference

Uncertainty estimation integrates three complementary principles:

1. Distribution Plausibility — evaluates whether predictions follow expected value ranges
2. Sampling Stability — measures consistency under stochastic decoding
3. Cross-field Consistency — enforces logical relationships across extracted variables

---

Citation

If you use this work, please cite:

@inproceedings{yu2026uncertainty,
  title={Uncertainty-Aware Structured Data Extraction from Full CMR Reports via Distilled LLMs},
  author={Yu, Yi and Martin, Parker and Bu, Zhenyu and Liu, Yixuan and Zheng, Yi-Yu and Simonetti, Orlando and Han, Yuchi and Xue, Yuan},
  booktitle={IEEE 23rd International Symposium on Biomedical Imaging (ISBI)},
  year={2026},
}