A Comprehensive Survey on Automated Radiology Report Generation: Methods, Explainability, Multimodal Alignment, and Clinical Integration

Abstract

Automating radiology report generation has become an important area of research because it can save reporting time and help maintain consistency in clinical diagnosis. In this survey, we reviewed recent papers that worked on different techniques for generating radiology reports. The approaches discussed in these papers include transformer models, multimodal learning methods that
connect images with text, contrastive learning frameworks, structured reporting formats, and radiology-specific large language models. Some works also used medical knowledge sources, lesion-based information, semantic tag prediction, dual stream encoders, and GPT-based text systems. We compared all studies based on their methods, datasets, evaluation metrics, strengths, and limitations. The major challenges identified include inaccurate or fabricated medical statements, weak multimodal reasoning, imbalance in datasets, lack of clinical testing, and difficulties in integrating these systems into real hospital workflows. Overall, this survey summarizes the current developments in radiology report generation and highlights the areas where improvement is still required so that these systems can become more reliable and useful in real clinical practice