CURE: Curriculum-guided Multi-task Training for Reliable Anatomy Grounded Report Generation

CURE is a curriculum learning framework that improves medical vision-language models' ability to generate accurate radiology reports with better visual grounding. The method achieves significant gains in grounding accuracy (+0.35 IoU), report quality (+0.192 CXRFEScore), and hallucination reduction (18.6%) without requiring additional training data.

CURE addresses a critical limitation in medical AI systems: the disconnect between generated text descriptions and visual evidence in medical imaging. While vision-language models have shown promise in automating radiology report generation, they frequently produce hallucinations and misaligned findings—a serious problem in clinical settings where accuracy directly impacts patient care. The framework tackles this through an error-aware curriculum approach that dynamically prioritizes difficult samples during training, forcing the model to learn spatial-textual alignment more rigorously.

The advancement builds on established machine learning principles but applies them specifically to medical imaging challenges. Curriculum learning has proven effective in various domains, but its application to grounded medical report generation represents meaningful progress in making AI systems more interpretable and clinically reliable. By leveraging publicly available datasets rather than requiring expensive new annotated data, CURE offers a practical pathway for improving existing models.

The quantified improvements are substantial enough to matter clinically. An 18.6% reduction in hallucinations directly translates to fewer false or unsupported findings in generated reports. The IoU gains demonstrate genuinely better spatial understanding, while CXRFEScore improvements indicate higher-quality overall report generation. For healthcare institutions and researchers, this approach enables safer deployment of automated report systems.

Looking forward, the open-source release of code and model weights enables broad adoption and further refinement. The framework's data-efficient nature makes it particularly valuable for specialized medical domains where extensive labeled data remains scarce. Future work likely involves validating CURE on other medical imaging modalities and integrating it into clinical workflows.

• →CURE improves radiology report grounding accuracy by 0.35 IoU through curriculum-guided multi-task training
• →The framework reduces hallucinations by 18.6% without requiring additional annotated data
• →Method works by dynamically adjusting training emphasis on harder samples to strengthen spatial-textual alignment
• →Open-source release enables practical adoption by healthcare institutions and researchers
• →Represents significant progress in making medical AI systems more interpretable and clinically reliable