Generating radiology reports from medical images is a labour-intensive task, exacerbated by the global shortage of radiologists and the complexity of image interpretation. Automatic radiology report generation (ARRG) has emerged to address this challenge, with AI-driven methods showing promise in accelerating diagnosis and improving workflow efficiency. While previous reviews primarily focus on conventional AI and deep learning models and often adopt non-systematic methodologies, this work provides a systematic PRISMA-based review of 51 studies published between 2015 and 2025, offering a rigorous synthesis of model architectures, learning paradigms, data strategies, and evaluation metrics. We categorize approaches from classical CNN-RNN frameworks to emerging paradigms including knowledge graphs, large language models (LLMs), retrieval-augmented generation (RAG), AI agents, and multimodal vision-language models. Our analysis highlights key challenges in clinical translation, including model explainability, factual accuracy, generalization across modalities, and deployment feasibility. We also critically evaluate benchmark datasets and performance metrics, identifying inconsistencies that hinder fair comparison and clinical adoption. Finally, we propose actionable research directions to advance clinically reliable, explainable, and generalizable ARRG systems. By linking technical developments to real-world clinical requirements, this review serves as a forward-looking reference for researchers and practitioners aiming to bridge the gap from AI prototypes to trusted clinical tools.
Hamma et al. (Mon,) studied this question.