Kwon, Jangho Choi, Kihwan 2024-01-19T09:37:55Z 2024-01-19T09:37:55Z 2022-03-07 2020 2375-933X https://pubs.kist.re.kr/handle/201004/113858 This study proposes a trainable multi-contrast windowing method in order to optimally choose contrast windows for deep learning-based CT segmentation. Existing contrast windowing methods use parameters predefined by radiologists or manufacturers. These predefined contrast windows, however, have not been proven to be optimal set for machine learning based approaches. We therefore propose a trainable multi-contrast windowing module which can be easily integrated into deep convolutional neural networks. For performance evaluation, we investigate the effects of the trainable multi-contrast windows by applying the proposed windowing modules to a deep learning based segmentation network measuring liver tumors. The results show significant performance improvement when the windowing parameters are trainable. The proposed method enhances the performance for medical image analyses compared to rule-based windowing methods. English IEEE Trainable Multi-Contrast Windowing for Liver CT Segmentation Conference 10.1109/BigComp48618.2020.00-80 1 IEEE International Conference on Big Data and Smart Computing (BigComp), pp.169 - 172 IEEE International Conference on Big Data and Smart Computing (BigComp) 169 172 US Busan, SOUTH KOREA 2020-02-19 2020 IEEE INTERNATIONAL CONFERENCE ON BIG DATA AND SMART COMPUTING (BIGCOMP 2020) 000569987500030 2-s2.0-85084365496 Proceedings Paper