Related Experiment Video
Updated: Jan 12, 2026
Standardized Data Acquisition for Neuromelanin-Sensitive Magnetic Resonance Imaging of the Substantia Nigra
Published on: September 8, 2021
Comparative diagnostic performance of artificial intelligence models in structural MRI for schizophrenia: A
Martin Kotochinsky1, Pandora Eloa Oliveira Fonseca2, Veronica Ramirez Lopera3
1Department of Medicine, National University of Cuyo, Argentina.
Introduction:
Timely diagnosis of schizophrenia is essential to ensure prompt treatment initiation and adherence. Structural magnetic resonance imaging (sMRI), when combined with artificial intelligence (AI), offers a promising avenue to enhance diagnostic accuracy. However, its performance and clinical use is a matter of debate.
Methods:
PubMed, Embase, and Cochrane databases were searched for studies using AI models with sMRI to diagnose schizophrenia in adults. Eligible models encompass traditional machine learning methods and deep learning (DL) architectures, utilizing diverse neuroanatomical inputs, including gray matter (GM) features and whole-brain (WB) structural data. The outcomes of interest were diagnostic performance metrics as: sensitivity (SE), specificity (SP), area under the curve (AUC).
Results:
A total of 16 studies were included, comprising 3601 participants. Overall pooled SE and SP were 0.76 (95 % CI: 0.71-0.80) and 0.78 (95 % CI: 0.73-0.82), respectively. When compared, DL models outperformed Support Vector Machine (SVM), achieving higher SP of 0.83 (95 % CI: 0.80-0.86) vs. 0.78 (95 % CI: 0.72-0.83), and AUC of 0.892 (95 % CI: 0.81-0.90) vs. 0.782 (95 % CI: 0.70-0.82). WB input models also outperformed GM performance, with SP of 0.86 (95 % CI: 0.78-0.92) vs. 0.80 (95 % CI: 0.73-0.85), and AUC of 0.89 (95 % CI: 0.70-0.93) vs. 0.816 (95 % CI: 0.71-0.84).
Conclusion:
AI models using sMRI show promising but provisional diagnostic performance for schizophrenia. Across studies, DL architectures and WB inputs generally achieved higher specificity and AUC than SVM and GM features. Prospective, multi-site external validation cohorts are needed before routine clinical implementation.
