Şeref Barbaros ARIK, Murat ŞİMŞEK, Aslıhan BARAN, Gökhan ŞAHİN
Turkish Journal of Neurology - 2026;32(3):237-244
Objectives: This study aimed to evaluate the diagnostic performance of an artificial intelligence (AI)-based decision support model that estimates the probability of dementia using only routine fluid-attenuated inversion recovery (FLAIR) magnetic resonance imaging (MRI) sequences and to assess its potential role as a triage tool in outpatient clinical settings. Patients and methods: This retrospective single-center study included 96 patients who underwent brain MRI and had established clinical and neuropsychological diagnoses between January 2021 and December 2024. A previously published convolutional neural network model trained on 6,400 open-access MRI scans was applied to our institutional dataset to evaluate its diagnostic performance in a real-world clinical setting. For clinical interpretability, a predefined threshold of 50% was used to classify cases, and receiver operating characteristic analysis was performed to assess discriminative performance. Results: The dementia group (n = 47) had a mean age of 71.76 +/- 8.32 years (range, 55 to 90 years) and included 25 females (53.2%) and 22 males (46.8%). The control (nondementia) group (n = 49) had mean age of 75.18 +/- 7.48 years (range, 55 to 89 years) and included 25 females (51.0%) and 24 males (49.0%). The AI model achieved an accuracy of 89.6%, sensitivity of 78.7%, specificity of 100%, and an AUC of 0.937. The dementia probability (%) strongly correlated with clinical diagnoses (r = 0.762, p < 0.001). Notably, no false-positive cases were observed in this cohort, indicating high reliability of positive classifications. McNemar and Binomial tests confirmed that the classification performance was significantly better than chance (p < 0.001). Conclusion: Artificial intelligence-derived dementia probability from routine FLAIR MRI provides a rapid and standardized output that may support clinical decision-making at the initial point of care. The absence of false-positive cases is consistent with the potential utility of the model as a rule-in decision-support tool, supporting prioritization of patients for further diagnostic evaluation. These findings highlight the potential of a simple, single-sequence AI approach to enhance workflow efficiency in neuroradiology practice without replacing clinical judgment.
