부산대학교 도서관

Machine learning approach to medical image segmentation becomes more prevalent in radiology. However, the performance of segmentation models is not yet sufficient high for practical applications in clinics. A key cause to the performance limitation is the lack of the valid medical images in a training set. A segmentation model trained with invalid medical images is prone to generate false segmentations. A feasible solution to remedy the performance problem is to purify medical images of the training set prior to generating the segmentation model. In this paper, we present practical and effective methods for purifying the medical images in CT/MRI scans. We utilize Mask R-CNN models in the purification methods along with effective software tactics. Our experiments show that the segmentation model trained with purified medical images yields an average of 16% performance improvement.