A Self-guided Framework for Radiology Report Generation

Automatic radiology report generation is essential to computer-aided diagnosis. Through the success of image captioning, medical has been achievable. However, lack annotated disease labels still bottleneck this area. In addition, image-text data bias problem and complex sentences make it more difficult generate accurate reports. To address these gaps, we present a self-guided framework (SGF), suite unsupervised supervised deep learning methods mimic process human writing. detail, our obtains domain knowledge from reports without extra guides itself extract fined-grain visual features associated with text. Moreover, SGF successfully improves accuracy length by incorporating similarity comparison mechanism that imitates self-improvement through comparative practice. Extensive experiments demonstrate utility in majority cases, showing its superior performance over state-of-the-art methods. Our results highlight capacity proposed distinguish fined-grained details between words verify advantage generating