Heatmaps autoencoders robustly capture Alzheimer’s disease’s brain alterations

Deep neural networks have achieved unprecedented success in diagnosing patients with Alzheimer's Disease from their MRI scans. Unfortunately, the decisions taken by these complex nonlinear architectures are difficult to interpret. Heatmap methods were introduced to visualize the deep learning models that are trained after classifying groups, but very few quantitative comparisons have been conducted so far to determine what approaches would be the most accurate in representing patterns learned by a deep learning model. In this work, we propose to use autoencoders to fuse the maps generated by different heatmap methods to produce a more reliable brain map. We establish that combining the heatmaps produced by Layer-wise Relevance Propagation, Integrated Gradients, and the Guided Grad-CAM method for a CNN trained using 502 T1 MRI scans provided by the Alzheimer's Disease Neuroimaging Initiative produces brain maps better capturing the Alzheimer's Disease effects reported in a large independent meta-analysis combining 77 voxel-based morphometry studies. These results suggest that our nonlinear maps fusion is a promising approach to take advantage of the great variety of heatmap methods recently published and produce a map with robust feature representation and less noise.