Automatic Generation of Pulmonary Radiology Reports with Semantic Tags

Authors:

Abstract
The chest X-ray is widely used in clinical practice for diagnosis and treatment. Among all chest diseases, pulmonary disease accounts for the majority, and the description of the lungs in the radiology report is the most complicated. According to this situation, this paper proposes a pulmonary radiology report generation model (SRN+BC-LSTM) based on semantic tags of radiograph. Firstly, for the problem that the image features extracted by CNN do not contain obvious semantic information, this paper selects the high frequency words related to abnormalities in the pulmonary radiology report as semantic tags, and trains the multi-label classifier. Secondly, a binary classifier is combined to improve the BLEU of generated normal reports since the Chinese radiology reports have roughly the same description for the normal samples. The experiment results indicated that our model is 13% higher than the baseline model in BLUE4.