Deep learning interpretability in neuroimaging: A comprehensive survey and methodological recommendations

Deep learning (DL) models have experienced a surge in popularity due to their capacity to directly learn from raw data in an end-to-end paradigm without relying on a separate feature extraction process that may be based on restrictive assumptions. The neuroimaging community has enthusiastically embraced DL as it strives to learn biomarkers from complex, multivariate, multimodal datasets. However, a broad replacement of human intelligence with DL in clinical environments is yet far from realization. One of the major obstacles to this transition is the opacity of DL models. A deep understanding of models is essential for their effective deployment in safety-critical domains such as healthcare, where transparency and trust hold substantial significance. We provide a comprehensive review of the interpretability literature, specifically focusing on the current status of DL interpretability in neuroimaging studies. Ultimately, we highlight strategies and insights necessary for successfully integrating DL technology in characterizing and addressing mental disorders.