Research: Evaluating Vision-Language Foundation Models for Automated Radiology Report G...

Research question: How effectively can large vision-language models generate clinically accurate and contextually relevant radiology reports from chest X-ray images compared to existing baselines?

Background & Motivation: Automated report generation from medical images has the potential to reduce radiologist workload and increase diagnostic consistency, especially in resource-constrained settings. Vision-language foundation models (such as CLIP, Flamingo, or LLaVA) have recently shown promise in bridging image understanding and natural language generation across domains.

Research Gap: Despite progress, the fidelity, clinical accuracy, and contextual appropriateness of reports generated by such models in real-world medical imaging scenarios remain under-explored. Most prior studies focus on small-scale or generic datasets, often neglecting rigorous evaluation against radiological standards and expert benchmarks.

Approach & Expected Contribution: This project will systematically assess state-of-the-art vision-language models for chest X-ray report generation using the MIMIC-CXR dataset. The methodology will involve fine-tuning existing models, establishing baseline comparisons, and evaluating outputs with both automatic metrics (BLEU, ROUGE, CheXbert) and expert human review. The study aims to identify strengths, limitations, and specific failure modes of current models in clinical contexts.

Significance: Results will provide actionable insights into the current capabilities and limitations of foundation models in clinical report generation, informing both model development and deployment considerations for AI-assisted radiology.