With this work, we provide a 3D whole brain segmentation models that outperforms the 2.5D approach FastSurferCNN. Furthermore, we explore the efficacy of spatial ensembling and its parameters. This resource contains the model files associated with final results of "Are 2.5D approaches superior to 3D deep networks in whole brain segmentation?" by Roy, Saikat; Kügler, David; Reuter, Martin to be published at the MIDL 2022 conference https://openreview.net/forum?id=Ob62JPB_CDF. These model files allow reproduction of our evaluation results and application of our models to other datasets. The accompanying source code is provided at https://github.com/Deep-MI/3d-neuro-seg .

unensembled_model: Weights of the 3D architecture trained without "Spatial Ensembling" labeled "Ours (w/o SL)" in the paper. ensembled_model_: Weights of the 3D architecture with all methodological options enabled labeled "Ours" in the paper. The index identifies the associated image region (image quadrant). Find further documentation w.r.t. image regions in model/QuadNet.py at https://github.com/Deep-MI/3d-neuro-seg (especially assign_nets_to_coords() )