Fig. 1.

Deep Learning pipeline. (A) Number of raw MRI datasets employed in the analysis before data augmentation, with and without matching demographic variables (sex, age, and education) for the people with bvFTD, AD and the HC. (B) Data preparation and augmentation pipeline consisting of random volume rotations, random flipping, Gaussian noise addition, volume scaling and enhancing by a zoom transformation. This set of augmentations increased the sample size by a factor of 10. (C) 3D DenseNet network architecture consisting of a sequence of dense blocks and transition layers consisting of a Batch Normalization (BN), a rectified linear unit (ReLU), and a convolution transformation, ending in a prediction layer to produce the output. (D) Model evaluation interpretation, with the performance metrics consisting of the ROC curve and an AUC report, a radar plot showing the accuracy, sensitivity, specificity, precision, recall, and F1 metrics. An occlusion sensitivity analysis to obtain the most relevant parts of the images for the classification, and a test subsample variability analysis to assess sample heterogeneity. AD: Alzheimer's disease; BvFTD: behavioral-variant frontotemporal dementia; HC: healthy controls.