ScaDO Net
He Wang, Weiwei Zhang · 2019
Segmentation of neuroanatomy based on fully convolutional networks (FCNs) has been proven to be powerful in numerous medical applications with excellent performance. However, the lack of annotated data and the inconsistency between different datasets make it difficult in training a deep network model for brain segmentation. In this paper, we propose a 3D patch-wise based model to learn the brain segmentation, namely ScaDO net (Scaffold-Dense-Octave net). This scaffold-like net is designed by Scaffold blocks to perform the 3D patch-wise brain segmentation task. Even with more than one hundred convolution layers, the number of parameters and FLOPs (Float Point Operations) are strictly controlled by a set of well-defined hyperparameters. Furthermore, the octave convolution is incorporated into the ScaDO net to reduce the FLOPs and keep the number of parameters. To give reasonable suggestions in practical applications, we have conducted ablation studies about FLOPs-DICE trade-off, which showed that setting appropriate hyperparameters would achieve a high DICE value with relatively small FLOPs. Compared with the state-of-the-art methods, the experimental results on different heterogeneous brain datasets have demonstrated that the proposed ScaDO net can achieve an accurate brain segmentation, even for small and delicate subcortical structures. With over one hundred convolution layers, the proposed method has as shorter than 40s for per MRI volume.