Generating Chinese Radiology Reports from X-Ray Images: A Public Dataset and an X-ray-to-Reports Generation Method

Deep learning methods have revolutionized medical image analysis, enabling tasks such as lesion classification, segmentation, and detection. However, these methods rely on annotations, posing a burden on healthcare professionals. In contrast, medical reports contain valuable information, leading to the emergence of Medical Reports Generation from Medical Images (MRGMI). Despite advancements, MRGMI predominantly focuses on English reports, lacking solutions for other languages. To address this and to generate responsible Chinese MRGMI model, we present a Chinese MRGMI dataset of over 40,000 Xray-image-report pairs, covering diverse diseases. We further provide 500 graphnode annotations of the reports and propose the CN-RadGraph model, extracting graph nodes from reports to, in a clinical-responsible way, evaluate our MRGMI model: Chinese X-ray-to-Reports Generation (CNX2RG) model. Considering linguistic disparities, we enhance the SOTA method with prompt training, graph-based augmentation, and sentence shuffling. Our CN-X2RG model shows significant improvements over baselines. The dataset and code are publicly available, fostering clinical-responsible research and development.