Automatic Chest Radiology Report Generation Using Reinforcement Learning

Kourav M.; Agarwal S.; Arya K.V.; Petrlik I.; Rodriguez C. · Lecture Notes in Networks and Systems (2023) · DOI: 10.1007/978-981-99-1912-3_22

Resumen

Radiology report generation is the manual task assigned to radiologists which analyze the given chest image (x-ray) of patient and generates the report based on that analysis. Many of the research is undergoing in this field to make this process automatic. However, there are a large number of difficulties which have been faced by the researchers while developing the efficient automatic report generation model. The main difficulties include the generation of clinically inaccurate reports when analyzed with the reports of the expert radiologist. The introduction of deep learning based approaches achieved tremendous advancement in these areas and can produce chest X-ray reports in a more accurate manner. BY further optimizing the deep learning based approaches, we proposed a reinforcement learning-based model that generates automated X-ray reports based on input chest X-ray images. We also developed a lookahead inference mechanism to overcome the limitation of beam search, which selects words based on policy and value networks at each particular step of time. The proposed model achieved 0.454, 0.546, 0.466, 0.410, and 0.424 for the Bleu-1, Bleu-2, Bleu-3, Bleu-4, and Rouge evaluation metrics, respectively, surpassing all available state-of-the-art algorithms.

Instituciones: Universidad Nacional Mayor de San Marcos; Universidad Nacional Federico Villarreal