Morphometric Explanations for Deep Learning Based Neuroimaging Classifiers

Despite increasing accuracy of deep-learning-based medical imaging classifiers for detecting complex neurodegenerative disorders, an inherent lack of transparency and interpretability in their decisions poses a challenge to their clinical and research applications. While existing explainability methods provide insights into broader aspects of a classification decision, they do not provide a direct interpretation of how specific image information was used for a classification. Geometric, or morphological, brain changes are known to exist in multiple neurodegenerative disorders, and explaining how these changes are being used by a classifier would provide domain experts with an easily interpretable understanding of morphology contributing towards the classification decision. Our work proposes a counterfactual method to explain morphological features used by deep-learning classifiers for classifying image instances. Furthermore, to provide a more interpretable understanding of common morphological features used by a classifier for the population, we combine the instance explanations to describe average morphological brain changes used by the classifier in MNI152 space. The proposed method was tested using classifiers trained for detecting Parkinson's Disease (PD) and a simulated disease in T1-w MRIs. The counterfactual images generated by our method were shown to explain learnt disease-related image information, as the classification of more than 98% of images for the PD classifier was changed by modifying the explained aspects of the image morphology. By deriving geometric explanations used by deep-learning classifiers, future research can advance trustworthiness by validating the classifiers against existing disease knowledge and potentially harness them to explore novel disease processes. Our code is available at: https://github.com/ankusharudkar/Morphometric-Explanations.