Search

Deep learning applied to MRI for Alzheimer’s classification is hypothesized to improve if the deep learning model implicates disease’s pathophysiology. The challenge in testing this hypothesis is that large-scale data are required to train this type of model. Here, we overcome this challenge by using a novel data augmentation strategy and show that our MRI-based deep learning model classifies Alzheimer’s dementia with high accuracy. Moreover, a class activation map was found dominated by signal from the hippocampal formation, a site where Alzheimer’s pathophysiology begins. Next, we tested the model’s performance in prodromal Alzheimer’s when patients present with mild cognitive impairment (MCI). We retroactively dichotomized a large cohort of MCI patients who were followed for up to 10 years into those with and without prodromal Alzheimer’s at baseline and used the dementia-derived model to generate individual ‘deep learning MRI’ scores. We compared the two groups on these scores, and on other biomarkers of amyloid pathology, tau pathology, and neurodegeneration. The deep learning MRI scores outperformed nearly all other biomarkers, including—unexpectedly—biomarkers of amyloid or tau pathology, in classifying prodromal disease and in predicting clinical progression. Providing a mechanistic explanation, the deep learning MRI scores were found to be linked to regional tau pathology, through investigations using cross-sectional, longitudinal, premortem and postmortem data. Our findings validate that a disease’s known pathophysiology can improve the design and performance of deep learning models. Moreover, by showing that deep learning can extract useful biomarker information from conventional MRIs, the advantages of this model extend practically, potentially reducing patient burden, risk, and cost.

Deep learning techniques on MRI scans have demonstrated great potential to improve the diagnosis of neurological diseases. Here, we investigate the application of 3D deep convolutional neural networks (CNNs) for classifying Alzheimer’s disease (AD) based on structural MRI data. In particular, we take on two challenges that are under-explored in the literature on deep learning for neuroimaging. First deep neural networks typically require large-scale data that is not always available in medical studies. Therefore, we explore the use of including longitudinal scans in classification studies, greatly increasing the amount of data for training and improving the generalization performance of our classifiers. Moreover, previous studies applying deep learning to classifying Alzheimer’s disease from neuroimaging have typically addressed classification based on whole brain volumes but stopped short of performing in-depth regional analyses to localize the most predictive areas. Additionally, we show a deep net trained to distinguish between AD and cognitively normal subjects can be applied to classify mild cognitive impairment patients, with classification scores aligning empirically with the likelihood of progression to AD. Our initial results demonstrate both that we can classify AD with an area under the receiver operator characteristic curve (AUROC) of .990 and that we can predict conversion to AD among patients in the MCI subgroup with an AURUC of 0.787. We then localize the predictive regions, by performing both saliency-based interpretation and rigorous slice and lobar level ablation studies. Interestingly, our regional analyses identified the hippocampal formation, including the entorhinal cortex, to be the most predictive region for our models. This finding adds evidence that the hippocampal formation is an anatomical seat of AD and a prominent feature in its diagnosis. Together, the results of this study further demonstrate the potential of deep learning to impact AD classification and to identify AD’s structural neuroimaging signatures. The proposed classification and regional analyses methods constitute a general framework that can easily be applied to other disorders and imaging modalities. Acronyms MRI: magnetic resonance imaging AD: Alzheimer’s disease CN: cognitively normal MCI: mild cognitive impairment MCIs: mild cognitive impairment stable MCIp: mild cognitive impairment progression HP: hippocampal formation ADNI: Alzheimer’s Disease Neuroimaging Initiative HCP: Human Connectome Project CNNs: convolutional neural networks CAM: class activation mapping ReLU: rectified linear unit BN: batch normalization AUC: area under the curve ROC: receiver operating characteristic AUROC: area under the receiver operating characteristic curve