الفهرس | Only 14 pages are availabe for public view |
Abstract Writing radiology reports in hospitals is a time-consuming activity that also necessitates competence from the radiologists involved. This thesis offers a deep learning model for generating radiological reports automatically based on a chest x-ray or a mammogram. There are three steps to our work: (1) Fine-tune a Chexnet that has already been trained to predict certain tags from an image. (2) Calculate weighted semantic features from the pretrained embeddings of the predicted tag. (3) Condition a pre-trained GPT2 model on the visual and semantic features to construct the comprehensive medical reports. The model was trained on the publicly available IU-XRay dataset that contains chest X-ray images and their corresponding reports, and on the CDD-CESM dataset that contains mammograms and their reports. The CDD-CESM dataset is a new dataset that we collected from Egypt’s National Institute of Cancer as no publicly-available dataset exists for mammograms with full-text report |