Exploring Alzheimer's anatomical patterns through convolutional networks

摘要

This work demonstrates the usage of Convolutional Neural Networks (CNNs) to explore and identify the brain regions most contributing to Alzheimer's disease in two-dimensional images extracted from structural magnetic resonance (MRI) images. In a first stage, we set up different CNN configurations which are trained in a supervised mode reaching classification accuracy similar to that in other works. Then, the best performing CNN is chosen and we create brain models for each filter at the CNN first layer as they convolve throughout MRI images of patient cases. The brain models are further explored as their corresponding filter activations throughout brain regions reveals different anatomical patterns for different patient class, and thus, allowing us to identify the CNN filters with greatest discriminating power and which brain regions contribute most. Specifically, the CNN shows the largest differentiation between patients in the frontal pole region, which is known to host intellectual deficits related to the disease. This shows how CNNs could be used to provide interpretability on Alzheimer's and constitute an additional tool to support decision making in clinical practice.