Learning like a radiologist: a medical vision-language model for radiological image analysis via curriculum learning

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Deep Learning for Healthcare

Background:

• Medical vision-language models (MVLMs) show potential but struggle with generalizability, data quality, and clinical relevance.
• Current MVLMs lack robust evaluation frameworks for complex diagnostic reasoning.

Purpose of the Study:

• To develop and evaluate RadiSim-CL, an MVLM trained with curriculum learning simulating a radiologist's three-phase learning pathway.
• To introduce RadiSim, a large-scale dataset of 12 million image-text pairs for MVLM training.
• To establish a comprehensive five-stage validation framework for evaluating MVLM performance in medical imaging.

Main Methods:

• Curriculum learning approach simulating foundational, anatomical, and diagnostic reasoning phases.
• Curated RadiSim dataset (12M image-text pairs) aligned with learning phases.
• Five-stage coarse-to-fine validation framework across 24 zero-shot subtasks (MR, CT, DR imaging).

Main Results:

• RadiSim-CL matches state-of-the-art performance on foundational and anatomical tasks.
• Demonstrated superior performance in complex reasoning: AUC 0.953 for brain tumor diagnosis, 0.764 accuracy for meningioma grading.
• Ablation studies confirmed the effectiveness of the curriculum learning strategy.

Conclusions:

• RadiSim-CL provides a scalable and clinically aligned solution for improving diagnostic accuracy in medical imaging.
• The curriculum learning approach and comprehensive evaluation framework advance MVLM development for clinical applications.