In the same vein, the way #CorridorKey was trained using synthetic data reminded me of the technique of using labeled CT scans to create a bunch of digitally rendered radiographs for training a machine learning model to detect regions of interest in 2D thorax xrays.

So I started wondering whether anyone had tried applying this to segmentation, and sure enough, #SynthSeg goes one step further and solves the inherent issue of the applicability of synthetic data by sampling from a generative model