Fig. 3. HIV diagnosis from MRIs.

a Age discrepancy (p = 0.0002, two-tailed two-sample t-test) between n = 223 control (Ctrl) subjects and n = 122 HIV patients resulted in the baseline ConvNet learning the confounding effects (b, d, f), which were alleviated by the proposed CF-Net (c, e, g). Boxplots are characterized by minimum, first quartile, median, third quartile, and maximum. b, c HIV-prediction scores measured on a subset of n = 122 control and n = 122 HIV subjects with the same age distribution (c-independent). d, e t-SNE visualization of the feature space learned by the deep-learning models. f, g Saliency maps³³ corresponding to the voxel-level attention (larger attention means more discriminative voxels) by the models.