Figure 3.

Scheme of evaluation of our CNN model in this study. We primarily used ADNI, AIBL, GENIC, and the UK Biobank to develop our model (turquoise scale). CNN1 works in a hold-out approach (data split: 80% train, 10% development, 10% test set, each set of the data indicated by arrows). CNN2 was trained as a 10-fold cross-validation model using the data of the four primary cohorts (turquoise scale) in the training loop. To evaluate the performance of these two models, we used AddNeuroMed (light blue) and J-ADNI (lilac) as external test sets. In CNN3, we added the external test sets in the 10-fold cross-validation. For comparison reasons, we evaluated our CNN2 scheme with skull-stripped T1w MRIs (CNN4). HO, hold-out; CV, cross-validation; ED, external test set; SS, skull-stripped images.