This repository tries to follow the architecure of Özgün Çiçek et al. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation as closely as possible. As seen in the below figure, the number of parameters in this implementation is very similar to those reported in the paper but not exactly the same. I think this difference may be due to the counting of the batch normalisation parameters.
There has been no attempt to optimise the performance of this architecture. However, a concise, modular, class-based implementation of the 3D U-Net is provided and as such it is probably more suited as a learning/teaching resource.