Abstract
A lightweight convolutional neural network model is presented for automatic brain tumor segmentation from multimodal MRI images. To explore 3D spatial contexts with computational efficiency, we introduce a Complementary Convolution Unit (CCU), which reduces memory overhead but not sacrifices accuracy by replacing 3D convolution with two complementary 2D convolutions. Specifically, 2D convolutions in different orientations are used to capture different in-slice features. Furthermore, we enhance the CCUs with dense connections to speed up network training and facilitate feature reuse. To mitigate interference between brain tissues, the task of multi-label brain tumor segmentation is decomposed into three binary segmentation subtasks. For each subtask, we fuse the segmentations on multi-axes to further improve the segmentation accuracy. Validations and comparisons with recent methods conducted on the BRATS17 validation dataset have demonstrated the effectiveness of proposed model. The experiment results showed that we achieved average Dice scores of 0.890, 0.808 and 0.753 for the whole tumor, tumor core and enhancing tumor core, respectively.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have