AbstractBackgroundDeep learning shows promise in detecting Alzheimer’s Disease (AD) based on brain MRI1. Models trained on limited samples may fail when applied to new datasets (other cohorts, scanners), but privacy concerns make it difficult to centralize large amounts of data2. To address this, we developed a general federated learning architecture3, which we applied to train a 3D convolutional neural network (CNN) to detect AD from T1‐weighted brain MRI data from the three independently collected phases of the Alzheimer’s Disease Neuroimaging Initiative (ADNI).MethodWe trained a 3D‐CNN over a federated learning environment of 3 sites corresponding to each ADNI phases (ADNI1, ADNI2/GO, ADNI3). No subject data are shared across sites, only model parameters, thus satisfying privacy/regulatory requirements. Each site trains the model on its own local dataset for several epochs and shares only the locally‐trained parameters with a central server. The server aggregates the parameters and computes the global model, which is sent back to each learner, and the process repeats.We trained each learner for 4 local epochs in between federation rounds for a total number of 25 rounds. We divided each ADNI dataset into training, validation and test sets with distribution 70/20/10% (Figure 1). We also compared the performance of training our CNN models with random initialization versus using pre‐trained weights from a UK Biobank MRI sex classification task.ResultTable 2 shows the accuracy, precision, recall and F1 scores for our federated CNN, compared to a centralized model that was trained over all ADNI data. All models achieved excellent performance for classifying AD (96‐98%). Overall, the federated model with pretrained weights performed best. Pre‐training boosted accuracy for the federated, but not the centralized, case.ConclusionFederated learning achieves high performance in AD prediction from MRI and promises new applications in neuroimaging.[1] Lu et al. A Practical Alzheimer Disease Classifier via Brain Imaging‐Based Deep Learning on 85,721 Samples. bioRxiv 2020_08_18_256594. 2021[2] Ming et al. COINSTAC: Decentralizing the future of brain imaging analysis. F1000Research 6:1512, 2017.[3] Stripelis et al. Scaling Neuroscience Research using Federated Learning. ISBI 2021.