Model Complexity of Deep Residual U-NET for CT Liver Volumetry

Abstract

Computed Tomography (CT) has been used for liver volume measurement because of the highest location accuracy. Automated segmentation methods may improve CT volumetry time, but it has low accuracy. Residual U-Net which is one of the deep learning methods could improve segmentation accuracy. However optimization of residual U-Net hasn’t been demonstrated yet. The purpose of this paper is to investigate the optimal complexity for CT liver volumetry. The study was conducted using the 3D-IRCADb01 Datasets (10 males, 10 females) published by MIS Training Center, 15 people learned and 5 people tested. Segmented images were generated using Deep Residual U-Nets with a total of four different complexity. As a result, as the model became more complex, the total parameters and training time increased exponentially. In all models, both training and testing showed more than 97% accuracy. All losses were less than 0.2. In the case of DCL, it was the lowest at 0.8037 in 3-layer and the highest at 0.9533 in 5-layer. In conclusion, 5 hidden layers of residual U-Net has the highest dice coefficient loss and could train the datasets faster than other complex models.