Can Anatomical Information Guide the Performance of Convolutional Neural Networks for Classifying Neurodegenerative Diseases Using Brain MRI?

Convolutional neural networks (CNNs) have been gaining outstanding success in classifying neurodegenerative diseases using brain magnetic resonance imaging (MRI) images. Given the highly informative content of brain MRI and the black-box nature of CNNs, it can be challenging to disentangle the decision-making process and identify cerebral regions that are relevant for the classification. In this study, we aimed to investigate the behavior of a 3D CNN for the classification of multiple system atrophy (MSA), a rare atypical parkinsonian syndrome, by exploiting a priori anatomical information related to this neurodegenerative disorder. We considered three regions of interest (ROIs) affected by pathological changes as input to a 3D CNN in combination with the T1-weighted MRI images. By comparing different CNN implementations and introducing a specific module accounting for the anatomical information, we achieved high accuracy (78–89%) in discriminating MSA patients from healthy controls. Furthermore, performances varied depending on the ROI used, with lower sensitivity for MSA patients when considering bigger regions. Instead, modifying the CNN with the anatomical gate module, an anatomically oriented attention mechanism, improved the classification using smaller regions. These findings represent an encouraging starting point to exploit fully disease-specific a priori knowledge and enhance classification performance while supporting better interpretability and advancing deep learning-based aid-to-diagnosis tools in clinical practice.