Automatic Radiology Report Generation: Approaches and Insights

In today’s era, deep learning finds widespread application across diverse domains. Exploiting its potential in healthcare can ease the workload of radiologists, allowing them to prioritize critical cases and essential treatments rather than dedicating time to report writing. Automatic report generation in healthcare parallels image captioning tasks, merging computer vision (CV) with natural language processing (NLP). This paper introduces and discusses the importance of automated generation of radiology reports. It explores the systematic evolution of different models, including CNNs, RNNs, and Transformer-based approaches, used for this purpose. It also provides background information and related work. Additionally, the paper highlights competitions that are driving progress in the field, aiming to keep researchers informed about recent advancements and encourage further exploration. Moreover, it outlines publicly accessible datasets containing radiology images and potentially related reports, facilitating additional research efforts. The paper also explains various evaluation metrics like BLEU, METEOR, ROUGE, CIDEr, etc., before concluding. These metrics provide scores ranging from 0 to 1, where higher values indicate better accuracy in capturing linguistic features and semantic content.