Multikernel Capsule Network for Schizophrenia Identification.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Schizophrenia significantly impacts patient quality of life.
• Machine learning methods, including simple and complex approaches, are used for schizophrenia identification based on functional connectivity.
• Existing methods have limitations in simultaneous optimization or require large datasets.

Purpose of the Study:

• To propose a novel multikernel capsule network (MKCapsnet) for schizophrenia identification.
• To overcome limitations of existing machine learning methods in feature extraction and classification.
• To leverage brain anatomical structure for improved analysis of functional connectivity.

Main Methods:

• Developed a multikernel capsule network (MKCapsnet) incorporating brain anatomical structure.
• Utilized kernels matched to brain partition sizes for capturing interregional connectivities at varying scales.
• Introduced capsule dropout in the capsule layer to mitigate model overfitting.

Main Results:

• The proposed MKCapsnet demonstrated superior performance compared to state-of-the-art methods.
• Analysis included performance comparisons with varying parameters and illustration of the routing process.
• The study confirmed MKCapsnet's potential for accurate schizophrenia identification.

Conclusions:

• MKCapsnet is a promising method for schizophrenia identification.
• This study is the first to apply capsule neural networks to functional connectivity MRI data.
• The novel multikernel capsule structure considering brain anatomy offers a new avenue for understanding brain mechanisms and neurophysiological signal classification.