Figure 1
Schematic illustration of the ReSeg image processing pipeline consisting of two convolutional neural networks responsible for defining a bounding box around the input MRI volume (NetCrop) and performing subsequent whole-brain segmentation on the cropped volume (NetReSeg). NetCrop is trained to predict the coordinates of a specific vertex point (pi0, pj0, pk0) and the lengths of the edges along the i, j, and k axes (pdi, pdj, and pdk, respectively) of the bounding box circumscribing the brain in the input MRI volume. The target output vector [i0, j0, k0, di, dj, dk] for each image is computed from the FreeSurfer mask. The coordinates of the final bounding box were determined using the center point (ci, cj, ck) of the bounding box predicted by NetCrop and fixed lengths (borderi, borderj, borderk) that had been defined based on the morphometric characteristics of adult human brains. The starting and end points of this bounding box along the i, j, and k axes are denoted by is, js, ks and ie, je, ke, respectively. This bounding box was applied to the input MRI volume and, during training, to the corresponding FreeSurfer mask (denoted by green circles). Cropped input volumes and masks were used to train the segmentation network NetReSeg. During inference, only MRI volumes are cropped and segmented. The figure was created with diagrams.net. JGraph Ltd: diagrams.net (Version 15.2.9) [Software].
