Morphological fingerprinting: Identifying patients with first-episode schizophrenia using auto-encoded morphological...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Psychiatric Disorders

Background:

• Structural brain abnormalities are investigated in schizophrenia using statistical and machine learning methods.
• Optimal characterization of these abnormalities for classification remains an open challenge.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach for classifying schizophrenia based on structural brain features.
• To compare the performance of autoencoded features against conventional morphological features.

Main Methods:

• A convolutional 3D autoencoder was trained on segmented brains from healthy individuals.
• Autoencoded morphological patterns were generated for schizophrenia patients and controls.
• A classifier was built using automated machine learning with autoencoded features and validated internally and externally.

Main Results:

• The autoencoder achieved satisfactory input reconstruction.
• Classifiers using autoencoded features outperformed those using conventional features by approximately 10%, reaching 73.44% accuracy (0.8 AUC) internally and 71.85% accuracy (0.77 AUC) externally.
• The approach demonstrated potential for identifying schizophrenia patients.

Conclusions:

• Automatically learned features from segmented brains enhance schizophrenia classification compared to traditional methods.
• Further improvements are necessary to establish this method as a clinical diagnostic marker.
• This study provides insights into applying deep learning to psychiatric disorders.