Radiomics analysis of striatal FDOPA PET imaging in patients with psychosis

Purpose: Schizophrenia is a debilitating psychiatric disorder characterized by an abnormal dopamine production. Right now, there is a lack of quantifiable biomarkers applied in its diagnosis and no way other than empirically to determine the patient's response to standard antipsychotic treatment. Radiomics is an image analysis technique which allows to analyze patterns not recognizable by the human eye, that may be of help for the creation of more reliable biomarkers. Methods: Radiomics features were extracted from 18F-DOPA PET scans of the striatal area and compared between patients and controls first, responders to standard treatment and non-responders second. The dataset consisted in 141 healthy controls and 137 patients (71 responders, 64 non-responders). Features were extracted from the SUVr signal using MIRP. Reproducibility analysis was conducted on separate test-retest scans and an ICC = 0.80 was applied as a threshold on the computed features. The remaining features were grouped using hierarchical clustering based on Spearman correlation. ANOVA testing was conducted on the two groups. Results: 15 features were selected. Linear regression showed an influence of gender and age on most features in patients. No difference was found between patients and controls. 10 features were significantly different between responders and non-responders. The feature with the highest area under the curve was the joint maximum (AUC = 0.66). Stepwise logistic regression did not show any improved performance using the combined features. Discussion: This study seems to confirm the influence of gender and age on the development of the disease. No difference was found between controls and patients, but patients came from a dataset too heterogenous. The differences between responders and non-responders seem to highlight that in responders the dopamine production is higher and creates a more irregular signal. Joint maximum was able to differentiate between responders and non-responders better than the SUVr mean, which is what is currently used in clinical practice. Validation on an independent cohort and the use of more complex classification algorithms may improve the results. Conclusion: Radiomics features may be a support for the creation of a biomarker able to predict treatment response in psychotic patients.