Co-Teaching Learning from Noisy Labeled FMRI Data for Diagnostic Classification of Major Depression

Brain imaging provides an effective tool for identifying brain abnormalities, and functional Magnetic Resonance Imaging (fMRI) is widely used due to its non-invasive nature and high spatiotemporal resolution. However, low signal-to-noise ratio, large inter-subject differences, and high cost of manual labeling may lead to noisy labels in fMRI-based diagnostic classification of neuropsychiatric disorders. In this study, we developed a co-teaching learning framework to address the issue of noisy labels in the fMRI-based diagnostic classification of Major Depression Disorder (MDD), in which a sample selection strategy based on loss value of predicted labels and true labels was realized through the dual-network collaborative architecture, dynamically screening out samples that may have noisy labels. Then the screened samples were used to update the parameters of the two networks. Experimental results revealed that our proposed method could achieve accuracy improvements of 2.2~3.2% in the classification between MDD patients and normal controls, suggesting that our method is helpful to address the issue of noisy labels in image-based diagnostic classification of neuropsychiatric disorders.