Interpretation of Brain Morphology in Association to Alzheimer's Disease Dementia Classification Using Graph Convolutional Networks on Triangulated Meshes

Abstract

We propose a mesh-based technique to aid in the classification of Alzheimer's disease dementia (ADD) using mesh representations of the cortex and subcortical structures. Deep learning methods for classification tasks that utilize structural neuroimaging often require extensive learning parameters to optimize. Frequently, these approaches for automated medical diagnosis also lack visual interpretability for areas in the brain involved in making a diagnosis. This work: (a) analyzes brain shape using surface information of the cortex and subcortical structures, (b) proposes a residual learning framework for state-of-the-art graph convolutional networks which offer a significant reduction in learnable parameters, and (c) offers visual interpretability of the network via class-specific gradient information that localizes important regions of interest in our inputs. With our proposed method leveraging the use of cortical and subcortical surface information, we outperform other machine learning methods with a 96.35% testing accuracy for the ADD vs. healthy control problem. We confirm the validity of our model by observing its performance in a 25-trial Monte Carlo cross-validation. The generated visualization maps in our study show correspondences with current knowledge regarding the structural localization of pathological changes in the brain associated to dementia of the Alzheimer's type.

Keywords: Alzheimer’s disease classification; Graph convolutional networks; neural network interpretability; triangulated meshes.

Fig.1.

Cortical meshes from a randomly selected HC subject (blue) and meshes of the subcortical structures from a randomly selected ADD subject (yellow). Presented are lateral views (a-b) of the HC’s left hemisphere (LH) and right hemisphere (RH) cortical meshes respectively. Medial views of the ADD subject’s LH and RH subcortical structure meshes are also presented (c-d).

Fig.2.

Single ResBlock in the GCN architecture used in this study. Linear mapping of F_in to F_out channels is implemented using a convolutional layer, $*\mathcal{G}$. This is done to match the number of input features to the number of desired feature maps.

Fig.3.

Residual GCN used for the binary classification of ADD. In this study, max-pooling operations are used to downsample the vertex dimension by a factor of 2.

Fig.4.

Monte Carlo cross-validation accuracy results for GCN and baseline model architectures from [29] used on brain meshes.

Fig.5.

Average TP CAMs on the cortical template from [10, 11] (top) and subcortical structures from [3] (bottom) including: (a-b, e-f) lateral-medial views of the LH respectively, (c-d, g-h) medial-lateral views of the RH respectively.