A Deep Learning Approach for Psychosis Spectrum Label Noise Detection from Multimodal Neuroimaging Data.

Area of Science:

• Neuroimaging
• Psychiatric Disorders
• Machine Learning

Background:

• Understanding brain mechanisms in mental disorders is complex.
• Neuroimaging techniques offer insights but have modality limitations.
• Psychosis nosology challenges biomarker identification.

Purpose of the Study:

• To introduce a deep convolutional framework for classifying and identifying label noise in brain imaging data.
• To apply this framework to structural and functional MRI data from schizophrenia patients.
• To distinguish potentially noisy subjects and investigate subtypes based on noise levels.

Main Methods:

• Developed a deep convolutional framework for neuroimaging data analysis.
• Applied the framework to structural and functional MRI data from a schizophrenia dataset.
• Utilized cross-validation and introduced a noise criterion for subject evaluation.

Main Results:

• The model learned from resting-state functional MRI data showed higher performance and informativeness compared to structural MRI data.
• A noise criterion effectively distinguished potentially noisy subjects for each modality.
• Analysis of borderline subjects revealed potential subtypes with distinct resting-state static functional connectivity features.

Conclusions:

• Schizophrenia patients can be differentiated from healthy controls using neuroimaging data.
• Resting-state functional MRI data is more informative and contains less label noise than structural MRI data.
• The developed framework aids in identifying noisy data and exploring potential schizophrenia subtypes.