CHESTPHENOT Released: Efficient and Auditable Radiology Report Parsing Model

Summary:The Yukkai team introduces CHESTPHENOT, a compact language model (0.5-3B parameters) designed for radiology report parsing. The model jointly extracts finding labels, three-class status (present/absent/uncertain), and verbatim supporting evidence spans. Trained using hybrid supervision and lightweight GRPO refinement, CHESTPHENOT demonstrates competitive performance under distribution shifts and significantly outperforms the CheXbert model in cross-institution detection. This model provides a re

---

Key Breakthroughs

CHESTPHENOT, developed by the Yukkai team, is a compact language model tailored for radiology report parsing with the following key features:

• Joint Extraction Capability: It can simultaneously extract finding labels, three-class status (present/absent/uncertain), and verbatim supporting evidence spans.
• Efficient Training Techniques: Trained using hybrid CheXbert+72B silver supervision, followed by supervised fine-tuning and lightweight GRPO refinement.
• Cross-Institutional Adaptability: Demonstrates strong performance on cross-institutional datasets, significantly outperforming the existing CheXbert model.
• Auditability: Over 99% of the final evidence spans are locatable in the source report, and the 3B model achieves an auditable-F1 of 47.5, outperforming Qwen2.5-7B one-shot prompting.

Technical Highlights

1. Hybrid Supervision and Optimization: CHESTPHENOT leverages CheXbert's supervision signals and 72B parameter model silver data, enhanced with GRPO optimization to improve generalization and accuracy.
2. Cross-Institutional Performance: The model excels in cross-institutional datasets, showcasing its potential for real-world applications.
3. Auditability: The evidence spans generated by the model are highly traceable, ensuring the reliability and auditability of radiology report parsing.

Industry Impact

CHESTPHENOT offers a more efficient and reliable tool for radiology report parsing in the healthcare AI sector. Its independence from external APIs makes it advantageous in terms of data privacy and security. Additionally, its strong performance on cross-institutional datasets positions it well for applications in multi-institutional collaborations and large-scale data processing.

Developer Recommendations

• Application Scenarios: Ideal for healthcare AI applications requiring efficient and auditable radiology report parsing.
• Training Data: Consider further fine-tuning with domain-specific silver data to enhance performance on specific tasks.
• Optimization Directions: Future work could explore joint training with multimodal data (e.g., images and text) to further boost model performance.

---

Source: ArXiv NLP/LLM (cs.CL) (2026-09-30)

— END —

Tags: #Medical Imaging #Language Model #AI in Healthcare #Auditability #Cross-Institutional