Figure 2
From: Dense cellular segmentation for EM using 2D–3D neural network ensembles
Methods. (a) Diagram of the 2D–3D + 3 \(\times\) 3 \(\times\) 3 network architecture, the best design tested in this paper. A 1-channel 3D image is passed through the network to produce a 7-channel output prediction of per-voxel probability distributions over the 7 label classes. Boxes represent multidimensional arrays, and arrows represent operations between them. Number triplets along box tops are array spatial axis sizes in (z, y, x) order. Numbers along box sides are array channel axis sizes. (b) Illustration of initialization-dependent performance of trained segmentation networks, and exploiting it for ensembling. An image of the test cell and ground truth labels are compared with segmentations of the best 4 trained 2D–3D + 3 \(\times\) 3 \(\times\) 3 network instances and an ensemble formed from them. The ensemble improves \({\rm{MIoU}}^{(org)}\) by 7.1% over the best single network.
