A Separate 3D-SegNet Based on Priority Queue for Brain Tumor Segmentation

To exploit three-dimensional (3D) context information and improve 3D medical image semantic segmentation, we propose a separate 3D-SegNet (S3D-SegNet) that combines 2D-SegNet with 1D-SegNet for brain tumor segmentation challenge. First, 2D-SegNet which has 4 encoding layers and 4 decoding layers is utilized to gain the 2D-features in a slice. Second, one-dimensional (1D) features are reassembled from the 2D features in the z-axis. Third, to use context along the z-axis, the 1D-features are inputted into 1DSegNet with 3 encoding and 3 decoding layers and then are classified feature-wise. Besides, we proposed a priority queue based on the weight values of samples for training to focus on the difficult and rare sample. Experimental results show that the separate 3D-SegNet can obtain higher accurate than 2D-SegNet and the priority queue can increase the speed of training convergence.