A Separate 3D Convolutional Neural Network Architecture for 3D Medical Image Semantic Segmentation

To exploit three-dimensional (3D) context information and improve 3D medical image semantic segmentation, we propose a separate 3D (S3D) convolution neural network (CNN) architecture. First, a two-dimensional (2D) CNN is used to extract the 2D features of each slice in the xy-plane of 3D medical images. Second, one-dimensional (1D) features reassembled from the 2D features in the z-axis are input into a 1D-CNN and are then classified feature-wise. Analysis shows that S3D-CNN has lower time complexity, fewer parameters and less memory space requirements than other 3D-CNNs with a similar structure. As an example, we extend the deep convolutional encoder–decoder architecture (SegNet) to S3D-SegNet for brain tumor image segmentation. We also propose a method based on priority queues and the dice loss function to address the class imbalance for medical image segmentation. The experimental results show the following: (1) S3D-SegNet extended from SegNet can improve brain tumor image segmentation. (2) The proposed imbalance accommodation method can increase the speed of training convergence and reduce the negative impact of the imbalance. (3) S3D-SegNet with the proposed imbalance accommodation method offers performance comparable to that of some state-of-the-art 3D-CNNs and experts in brain tumor image segmentation.