Label Noise-Robust Ensemble Deep Multimodal Framework For Neuroimaging Data

Neuroimaging data have become widely studied in the context of identifying brain-based markers of mental illness. however, this work is hampered by the use of symptom and self-report assessments of diagnosis, as well as lack of clarity in the nosological categories. Hence, treating existing diagnostic categories as label noise problems might be beneficial. Ensemble methods and deep learning models were used in many applications and revealed remarkable findings dealing with label noise. In this study, we incorporated deep convolutional frameworks and bagging approaches for diagnostic classification, identifying potential biomarkers and mitigating the effects of label noise across mood and psychosis categories using structural and functional MRI data. We conducted repeated k-fold cross-validation techniques to train individual base models on different subsets of data and aggregate independent models for final classification. Moreover, we interpreted the results and identified class-specific relevant learned features contributing to a successful diagnosis and highlighted differences for different modalities. Overall, our proposed method shows improvement in classification performance.