Liver segmentation with 2.5D perpendicular UNets. (May 2021)

• Record Type: Journal Article
• Authors: Han, Lin; Chen, Yuanhao; Li, Jiaming; Zhong, Bowei; Lei, Yuzhu; Sun, Minghui
• Is Part Of: Computers & electrical engineering. Volume 91(2021)
• Publication Date: 2021-05
• Subjects: Liver Segmentation -- Res-UNet -- Deep Learning -- Model Fusion -- Medical Imaging; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2021.107118
• Languages: English
• ISSNs: 0045-7906

Abstract

Highlights: Liver and hepatic tumor segmentation are essential for computer aided diagnostics. Deep learning based segmentation methods are effective in medical image segmentation. Adequate preprocessing and postprocessing improves deep learning model's accuracy. Incorporating residual connections in UNet speeds up model convergence. Fusing results of multiple deep learning models improves segmentation accuracy. Abstract: Liver and hepatic tumor segmentation is a crucial yet challenging step during the screening and diagnosis of liver illnesses. Currently, accurate 3D segmentation deep learning models are large, while the smaller 2D ones are generally less accurate due to their small receptive fields. To reduce model sizes and increase segmentation accuracy, we propose 2.5D Perpendicular-UNet to fuse the segmentation results of three perpendicular 2.5D Res-UNets in the task of liver and hepatic tumor segmentation. Data augmentation, loss functions, and post-processing steps are customizable with our model. With a larger receptive field in three dimensions, our model outperforms 2D UNet models in accuracy, achieving 0.962 and 0.735 Dice scores for liver and tumor segmentation on the liver tumor segmentation dataset. Being smaller than 3D models, our 2.5D P-UNet trains using less data and GPU memory. This enables it to be deployed on low-configuration hardware, expanding its potential use. Graphical abstract: Image, graphical abstract