Abstract
Interpretability is essential to MRI brain studies relying on deep learning models for neuroscientific discovery. One way to facilitate the interpretability of a deep learning model is to ensure the samples are arranged in the model's latent space with respect to clinically meaningful variables. To achieve this in the context of cross-sectional brain MRI studies, we regularize the latent space of a multi-label classifier via pairwise disentanglement, so that the difference between the representation of two brain MRIs along the disentangled direction in the latent space is similar to the difference in their neuropsychological test scores. We apply our technique to classify brain MRIs of 156 controls, 165 cases diagnosed with mild cognitive impairment (MCI), 166 diagnosed with human immunodeficiency virus (HIV)-associated cognitive disorder (HAND), and 32 individuals diagnosed with HIV without HAND. The latent space is disentangled with respect to the neuropsychological z-score (NPZ), which is negatively correlated with the severity of cognitive impairment (i.e., low scores for those diagnosed with MCI or HAND). Based on cross-validation, the proposed model achieves statistically significantly higher balanced accuracy than the same model without disentanglement. Furthermore, the difference between representations along the disentangled direction significantly correlates with the difference in NPZ. Finally, the brain regions guiding the classification process aligned with the neuroscientific literature.