Chain-of-Thought Prompting with Retrieval-Augmented Generation for Explainable Chest Radiograph Diagnosis: A Review of Vision-Language Models

Abstract

The integration of multimodal large language models (LLMs) into medical image interpretation has introduced new possibilities for clinical decision support, particularly through chain-of-thought (CoT) prompting and retrieval-augmented generation (RAG). This review examines existing CoT prompting strategies and RAG-enhanced approaches that aim to improve both the diagnostic accuracy and clinical explainability of vision-language models (VLMs) in medical imaging analysis, with chest X-ray (CXR) interpretation as the primary application domain. A two-dimensional analytical framework is proposed to categorize current approaches along reasoning granularity and evidence grounding depth. Representative methods published between 2023 and 2025 are systematically reviewed, evaluated across public benchmarks including MIMIC-CXR, CheXpert, VQA-RAD, SLAKE, and PathVQA. The analysis indicates that structured multi-step CoT prompting combined with domain-specific retrieval consistently outperforms zero-shot and standard prompting baselines, with factual accuracy improvements ranging from 20.8% to 43.8% when RAG modules are incorporated. Six evaluation dimensions for explainability assessment in medical imaging contexts are identified. Limitations of current approaches and directions for future investigation are discussed.