Abstract
Automatic medical report generation based on deep learning can improve the efficiency of diagnosis and reduce costs. Although several automatic report generation algorithms have been proposed, there are still two main challenges in generating more detailed and accurate diagnostic reports: using multi-view images reasonably and integrating visual and semantic features of key lesions effectively. To overcome these challenges, we propose a novel automatic report generation approach. We first propose the Cross-View Attention Module to process and strengthen the multi-perspective features of medical images, using mean square error loss to unify the learning effect of fusing single-view and multi-view images. Then, we design the module Medical Visual-Semantic Long Short Term Memorys to integrate and record the visual and semantic temporal information of each diagnostic sentence, which enhances the multi-modal features to generate more accurate diagnostic sentences. Applied to the open-source Indiana University X-ray dataset, our model achieved an average improvement of 0.8% over the state-of-the-art (SOTA) model on six evaluation metrics. This demonstrates that our model is capable of generating more detailed and accurate diagnostic reports.