RL for Medical Imaging: Benchmark vs. Clinical Performance

This paper highlights a critical issue in applying Reinforcement Learning (RL) to medical imaging: optimization for benchmark performance can lead to a degradation in cross-dataset transferability and, consequently, clinical utility. The study, using a vision-language model called ChexReason, demonstrates that while RL improves performance on the training benchmark (CheXpert), it hurts performance on a different dataset (NIH). This suggests that the RL process, specifically GRPO, may be overfitting to the training data and learning features specific to that dataset, rather than generalizable medical knowledge. The paper's findings challenge the direct application of RL techniques, commonly used for LLMs, to medical imaging tasks, emphasizing the need for careful consideration of generalization and robustness in clinical settings. The paper also suggests that supervised fine-tuning might be a better approach for clinical deployment.

• •RL optimization for benchmarks can hurt cross-dataset generalization in medical imaging.
• •The study suggests that the RL paradigm, specifically GRPO, may be overfitting to the training data.
• •Supervised fine-tuning might be a better approach for clinical deployment requiring robustness.
• •Structured reasoning scaffolds offer minimal gain for medically pre-trained models.