Saved in this browser. Sign in to sync across devices.
Clinical distinction: Schizophrenia is evaluated through a comprehensive clinical assessment, including possible alternative explanations for symptoms. A scan or machine-learning score is not a stand-alone diagnostic test, and this review does not establish an FDA-authorized schizophrenia diagnostic AI system.
What is radiomics?
Radiomics extracts quantitative patterns from medical images such as MRI scans. Machine-learning models look for combinations of those patterns that may differ between groups with and without a recorded clinical diagnosis. Researchers hope to learn more about brain biology or eventually improve clinical decision support—not to replace the history, examination and professional evaluation required today.
REVIEW SIZE
14 studies
The authors reviewed diagnostic and treatment-response applications of radiomics. Eight studies informed the main pooled diagnostic discrimination analysis.
DIAGNOSTIC DISCRIMINATION
AUC 0.81
Pooled area under the curve was 0.81 (95% CI 0.73–0.88). This describes ranking performance across study samples, not certainty about any one person.
SENSITIVITY AND SPECIFICITY
About 74% each
Pooled sensitivity was 73.6% and specificity 73.5%, leaving substantial potential for false negatives and false positives.
BETWEEN-STUDY VARIATION
I² 84.2%
Diagnostic studies differed substantially in methods, samples, scanners and validation—limiting transfer to a new clinical population.
Does an AUC of 0.81 mean an 81% chance of diagnosis?
No. AUC summarizes how a model ranks study cases versus controls across many possible decision thresholds. It is not the probability that a particular person's image gives a correct diagnosis. Sensitivity and specificity also depend on study design; they do not directly tell a family whether the person has schizophrenia. Real-world positive and negative predictive values depend on who is being assessed and how common the condition is in that setting.
Can imaging predict which treatment will help?
The review found exploratory treatment-response discrimination, including longitudinal prediction AUC around 0.80 and cross-sectional AUC around 0.88. Those pooled statistics came from varied research designs. They do not demonstrate a reliable way to select an antipsychotic, predict a person's recovery or improve outcomes through AI-guided treatment in routine clinical practice.
What prevents clinical use right now?
The authors found considerable methodological heterogeneity and limited external validation. In studies with weaker validation procedures, model performance tended to look more optimistic. Models can inadvertently learn patterns related to a scanner, hospital, medication exposure or study selection rather than generalizable illness markers. Better prospective studies and independently collected datasets are essential.
Questions to ask about advertised AI diagnostic scans
- Was the system independently tested in different clinics and people with similar clinical questions?
- How often does it produce false alarms or miss people who need care?
- Is the exact clinical use authorized, and is there evidence it improves outcomes?
- How is consent, privacy and demographic variation in accuracy handled?
- Does a qualified clinician integrate findings into a comprehensive evaluation?
Useful science, not a shortcut to diagnosis
For now, people experiencing hallucinations, delusions or other concerning changes deserve timely professional assessment and evidence-based support. Imaging may sometimes be used to investigate alternative medical causes, but AI radiomics cannot replace that clinical process.
Original meta-analysis and clinical guidance
- Schizophrenia — Caiazza and colleagues, October 21, 2026 ↗
Original peer-reviewed systematic review and meta-analysis of 14 studies.
- NIMH — Understanding Psychosis ↗
Current clinical assessment and differential diagnostic guidance.
- NIMH — Schizophrenia ↗
Current treatment and clinical-care information.
Evidence checked October 22, 2026 · Educational editorial draft; independent clinical review pending. No AI diagnostic product is endorsed.