Improvement of Low-Dose CT diagnosis prediction based self-supervised denoising

Abstract

Self-supervised learning denoising methods improve image quality for disease diagnosis and downstream classification prediction with only original images without clean images. Recently, various methods have been proposed to incorporate deep learning techniques in medical image analysis, but medical images are challenging to deal with due to several significant causes. First, medical images such as CT and MRI have noise by the imaging device, which confuses the model in learning, negatively affecting the prediction performance. Second, the medical field does not provide a clean image corresponding to the ground truth required for supervised learning. In this paper, we propose a method to improve classification performance by learning the loss of the self-supervised denoising model and the loss of the classification network by combining them. Furthermore, fine-tuning improves the classification prediction performance of a small number of medical image datasets.