A COMPARATIVE STUDY OF DEEP ARCHITECTURES FOR VOXEL SEGMENTATION IN VOLUME IMAGES

Abstract. This study investigates the performance of eight different deep learning architectures for voxel segmentation in volume images. The motivation is to segment carbon in carbon reinforced concrete (CRC) in micro-tomography (μ-CT) data. Although there are many 3D convolutional neural networks (CNNs) available, it is not yet clear which one works best for these specific tasks. In this study, the authors compare the following networks: DenseVoxNet, HighResNet, Med3D, Residual 3D U-Net, 3D SkipDenseSeg, 3D U-Net, V-Net, and LV-Net. To provide a more general recommendation for selecting a neural network, three medical datasets were added in addition to the three CRC datasets to facilitate the selection of an appropriate network based on the dataset characteristics. The experiments emphasize the importance of the initial random state, used for example to initialize the network weights. On average, the 3D U-Net is the best generalizing CNN, followed by the Residual 3D U-Net and the 3D SkipDenseSeg. While the 3D U-Net is a good architecture to start with, the experiments show that it does not perform best on all domains. To achieve optimal results, the authors propose recommendations for selecting a 3D neural network based on the dataset attributes.