Artificial neural network visualization methods reveal diagnostically relevant brain regions to detect Alzheimer’s disease: The first step towards comprehensive artificial intelligence

Abstract

AbstractBackgroundAlthough deep learning approaches achieve high diagnostic accuracy for detecting Alzheimer’s disease (AD) based on MRI scans, they are rarely applied in clinical research due to a lack of suitable methods for model comprehensibility and interpretability. Recent advances in convolutional neural networks (CNN) visualization algorithms may help to overcome these problems.MethodWe implemented a CNN model structure, trained it on 662 T1‐weighted MRI scans obtained from ADNI in a twentyfold cross‐validation procedure, and validated it on 1655 cases from three independent samples. Various CNN visualization algorithms were compared, which generated relevance maps indicating the contribution of individual image areas for detecting AD. We developed an interactive web application to display the 3D relevance maps and interactively change various visualization parameters. We assessed the clinical utility of relevance maps by comparing hippocampus relevance scores with hippocampus volume.ResultAcross the three independent validation datasets (Fig. 1), group separation showed high accuracy for AD dementia versus controls (AUC≥0.92) and moderate accuracy for amnestic mild cognitive impairment (MCI) versus controls (AUC≥0.73). Relevance maps obtained from Layer‐wise Relevance Propagation (LRP) provided a high spatial resolution and strongest focus (Fig. 2 & Fig. 3). LRP maps indicated that hippocampal atrophy was the most informative factor for AD detection (Fig. 4), with additional contributions from atrophy in other cortical and subcortical regions. Relevance scores within the hippocampus were highly correlated with hippocampal volumes (median: r=‐0.81, Fig. 5). The relevance maps of individual patients revealed additional clusters in the frontal and occipital lobe, which may be an indicator of model overfitting or potential bias of the training sample. When comparing the twenty cross‐validation models, stronger focus of the models on irrelevant areas was associated with lower accuracy to detect MCI.ConclusionThe relevance maps highlighted atrophy in regions that we had hypothesized a priori, which strengthens the overall comprehensibility and validity of the CNN models.