Exposure Fusion Framework in Deep Learning-Based Radiology Report Generator

Abstract

Writing a radiology report is time-consuming and requires experienced radiologists. Hence a technology that could generate an automatic report would be beneficial. The key problem in developing an automated report-generating system is providing a coherent predictive text. To accomplish this, it is important to ensure the image has good quality so that the model can learn the parts of the image in interpreting, especially in medical images that tend to be noise-prone in the acquisition process. This research uses the Exposure Fusion Framework method to enhance the quality of medical images to increase the model performance in producing coherent predictive text. The model used is an encoder-decoder with visual feature extraction using a pre- trained ChexNet, Bidirectional Encoder Representation from Transformer (BERT) embedding for text feature, and Long-short Term Memory (LSTM) as a decoder. The model’s performance with EFF enhancement obtained a 7% better result than without enhancement processing using an evaluation value of Bilingual Evaluation Understudy (BLEU) with n-gram 4. It can be concluded that using the enhancement method effectively increases the model’s performance.