Fig. 8

Our deep learning-based approach to subtyping. A General idea behind Smile-GANs (B) Schematic diagram of Smile-GANs. CN: healthy control, PT: patient, SUB: subtype (C) Network architectures of the three main functions (f, g, D): blue arrow represents one linear transformation followed by one leaky relu function, green arrow represents one linear transformation followed by one softmax function, red arrow represents only one linear transformation. The GAN network learns to synthesize MDD patient scans from scans of normal controls, contingent upon subtype, SUB, which is learned in the training process. The discriminator, D, of this GAN ensures that these synthesized scans are indistinguishable from real MDD patients scans. The mappings, f, estimated in this process capture features of the neuropathologic processes that transform imaging data of controls to those of MDD patients, adapted from Yang et al. [119]