부산대학교 도서관

Urinary stone is the most common abnormality occurring in human urinary system. Early detection in medical imaging is important for promptly treatment or surgery planning. Recently, deep learning methods have been proposed and played a major role in medical image applications. The performance of deep learning model largely depends on the size of training samples. However, the number of samples of patient having urinary stones is limited, and the process of labeling preparation is time-consuming and need experts. In this research, we used a combination of real stone-contained images and synthesized stone-embedded images to train U-Net model to segment urinary stones in abdominal x-ray images. Furthermore, the simple augmentation methods which appear realistic in x-ray images were also implemented in these combined datasets. From the comparative experiments, the proposed method can achieve higher F 2 score than U-Net trained with only real images. This method not only helps to reduce the requirement of training data and labels of positive samples, but also increase the number and variability of training data to improve deep learning performance.