Figure 2.

Overall architecture of the proposed approach. (1) CAE-based unsupervised learning is conducted to extract sparse representations; (2) the AD vs. NC classifier is constructed through supervised fine-tuning; (3) to construct the pMCI vs. sMCI classifier of pMCI vs. sMCI, the representations of the AD vs. NC task are considered as initial weights under a transfer learning manner. Finally, the important biomarkers are localized using a gradient-based approach.