Fig. 2

Architecture of the proposed 3D brain segmentation network (Brain SegNet) for brain lesion segmentation from MRIs. The input is multi-modality 3D MRI volume data. It has four convolutional blocks, and contains 17 convolutional layers in total, with residual units. It includes a refinement module capable of aggregating rich fine-scale 3D volume features over multiple convolutional blocks. An adaptive layer and an refinement layer are applied to each block for computing multi-level convolutional features