Label Noise-Robust Ensemble Deep Multimodal Framework For Neuroimaging Data.

Area of Science:

• Neuroimaging
• Computational Psychiatry
• Machine Learning

Background:

• Neuroimaging studies aim to find brain-based markers for mental illness.
• Current diagnostic methods rely on subjective symptom reports and unclear nosological categories, introducing label noise.
• Ensemble and deep learning methods show promise in handling label noise in various applications.

Purpose of the Study:

• To develop a robust diagnostic classification framework for mood and psychosis using neuroimaging data.
• To mitigate the impact of diagnostic label noise on classification accuracy.
• To identify potential neuroimaging biomarkers associated with specific diagnostic categories.

Main Methods:

• Incorporation of deep convolutional neural networks and bagging ensemble approaches.
• Utilizing structural and functional magnetic resonance imaging (MRI) data.
• Employing repeated k-fold cross-validation for model training and aggregation.

Main Results:

• The proposed method demonstrated improved classification performance for mood and psychosis categories.
• Class-specific relevant features contributing to diagnosis were identified.
• Differences in feature relevance across different MRI modalities (structural vs. functional) were highlighted.

Conclusions:

• The integration of deep learning with ensemble methods offers a promising strategy for improving diagnostic accuracy in psychiatry.
• This approach effectively addresses the challenge of label noise in neuroimaging-based mental illness classification.
• The identified biomarkers and modality-specific insights can aid in refining diagnostic criteria and understanding underlying neural mechanisms.