Feature and decision-level fusion for schizophrenia detection based on resting-state fMRI data

Area of Science:

• Neuroimaging
• Medical Diagnostics
• Machine Learning

Background:

• Early diagnosis of schizophrenia remains a significant clinical challenge.
• Resting-state functional magnetic resonance imaging (Rs-fMRI) offers potential for computer-aided diagnosis.
• Investigating advanced fusion techniques can enhance diagnostic performance.

Purpose of the Study:

• To evaluate decision-level and feature-level fusion schemes for discriminating schizophrenia from healthy subjects using Rs-fMRI data.
• To compare the efficacy of different Rs-fMRI feature types and fusion strategies.
• To optimize feature selection for improved diagnostic accuracy.

Main Methods:

• Rs-fMRI data preprocessing and denoising.
• Extraction and selection of four fMRI features: regional homogeneity, voxel-mirrored homotopic connectivity, fractional amplitude of low-frequency fluctuations, and amplitude of low-frequency fluctuations.
• Application of feature selection via concave minimization (FSV) and Support Vector Machine (SVM) classifiers on the COBRE dataset (70 schizophrenia, 70 healthy subjects).

Main Results:

• Decision-level fusion achieved 97.85% accuracy, 98.33% sensitivity, and 96.83% specificity.
• Feature-fusion scheme yielded 98.57% accuracy, 99.71% sensitivity, 97.66% specificity, and an AUC of 0.9984.
• Both fusion strategies significantly outperformed single-feature approaches.

Conclusions:

• Decision-level and feature-level fusion schemes substantially enhance the performance of schizophrenia detection based on Rs-fMRI.
• Fusion techniques represent a promising avenue for improving early and accurate diagnosis of schizophrenia.
• Rs-fMRI combined with advanced machine learning fusion methods offers a powerful tool for psychiatric disorder diagnosis.