Rethinking Medical Report Generation: Disease Revealing Enhancement with Knowledge Graph

Knowledge graph (KG) is an important component in medical report generation because it can reveal the relations among diseases and thus is often utilized during the generation process. However, the collection of a comprehensive KG is time-consuming and its usage is under-explored. In this paper, we construct a complete KG on chest images that includes most types of diseases and/or abnormalities.We further explore the usage of a KG in different directions. Firstly, by designing a rule-based criterion to classify disease types at sentence level, we find that long-tailed problems exist in the disease distribution and that generated reports from current advanced methods are far from being clinically useful. We alleviate the long-tailed distribution problem through a new augmentation strategy that increases the disease types in the tailed distribution. A two-stage generation approach based on image-level classification result is proposed in parallel to better capture ``disease-specific" information. On the other side, radiologists evaluate generated reports on whether they describe the diseases appearing in the input image. Following this idea, we propose diverse sensitivity (DS), a new metric that checks whether generated diseases match ground-truth and measures the diversity of all generated diseases. We observe that current leading methods cannot generate satisfying results and the proposed two-stage generation framework and augmentation strategies improve DS by a considerable margin.