Explicability in resting-state fMRI for gender classification

Artificial Intelligence, especially deep neural networks, have shown impressive performances for classification tasks since the last decade. In the medical field, trustworthy deep models exist but they do not provide any insights on how and why they classify data due to their complex structure. In this study we propose to leverage the power of deep neural network for classifying resting state brain activities by gender, then we use explainable Artificial Intelligence models to determine which functional networks are salient with respect to the gender. Firstly, we trained an accurate convolutional neural network to determine gender based on resting-state brain spatial maps corresponding to intrinsically connected networks and computed by independent component analysis. Then, we compare, through mask-based assessment, state of the art explainable Artificial Intelligence models to extract the most meaningful components involved in gender determination. Based on a powerful deep classifier, and with an appropriate explainable artificial intelligence method, we supply meaningful results in accordance with neurology literature results for gender classification. Throughout this study, we show that powerful deep models can be used in medical diagnostics since they recover, thank to reliable explainable artificial intelligence models, already established literature results related to gender determination with respect to brain network activities.