Fig. 1.
Diagram of the DeepAtrophy deep learning algorithm for quantifying progressive change in longitudinal MRI scans. During training, DeepAtrophy consists of two copies of the same “basic sub-network” (Dθ) with shared weights θ. Dθ is a 3D ResNet image classification network with 50 layers (Chen et al., 2019; He et al., 2015) and the output layer having k = 5 elements. Dθ takes as input two MRI scans from the same individual in arbitrary temporal order. The outputs from the two copies of Dθ feed into a 2k × m fully connected layer with weights ω. The resulting “super-network” Sθ,ω, takes as input two pairs of same-subject images, in arbitrary order, and with constraint that the inter-scan interval of one scan pair contains the inter-scan interval of the other scan pair. DeepAtrophy minimizes a weighted sum of two loss functions: the scan temporal order (STO) loss, which measures the ability of Dθ to correctly infer the temporal order of the two input scans; and the relative interscan interval (RISI) loss, which measures the ability of the super-network Sθ,ω, to infer which of the input scan pairs has a longer inter-scan interval. During testing, network Dθ is applied to pairs of same-subject scans. A single measure of disease progression, the predicted interscan interval (PII), is computed as a linear combination of the k outputs of Dθ. The coefficients of this linear combination are obtained by fitting a linear model on the subset of the training data (amyloid negative normal control group) with actual inter-scan interval as the dependent variable and outputs of Dθ as independent variables.