Consider inputting adjacent slices to train a multi-channel model

[jcohenadad]

Similar to 2.5D models, the idea is to consider adjacent slices as additional input channels for training and inference.

Implemented in afce2df, below is the result comparing single-slice estimation with 3-slice estimation:

singleslice_vs_3slice.mov

The 3-slice definitely looks more regularized along the rostro-caudal axis.

[jcohenadad]

Another way to look at it: Sagittal orientation, look at the regularity of the segmentation along the rostro-caudal direction, switching between pure 2D vs. 2.5D training: