Fig. 1
From: Generalizable deep learning framework for 3D medical image segmentation using limited training data
Inference workflow. The 3D image volume is split into three stacks of 2D images in transversal, sagittal, and coronal directions, respectively. The same 2D U-Net CNN performs inference on all three stacks of 2D images. This results in three image volumes. The final voxel-wise classification is achieved by averaging the probabilities