Diagnostic Accuracy of GPT-5 in Detecting and Localizing Acute Ischemic Stroke on Isolated Diffusion-Weighted Imaging: A Retrospective Case–Control Study

Abstract

Background/Objectives: Multimodal large language models are emerging tools for medical image interpretation, but their performance on isolated diffusion-weighted imaging (DWI) for acute ischemic stroke remains largely unexplored. We evaluated GPT-5 for stroke detection and vascular territory classification on preselected representative DWI images using a structured two-step prompting protocol. Methods: This single-center, retrospective, balanced case–control study included 220 cases (110 stroke patients and 110 stroke-negative controls). Four representative axial DWI slices per case were provided without apparent diffusion coefficient maps. Step 1 queried stroke presence and vascular territory; Step 2 queried the middle cerebral artery (MCA) segment-associated territory (M1–M4) following an MCA response. The reference standard was radiologist–neurologist consensus based on full multimodal magnetic resonance imaging, blinded to GPT-5 outputs. Diagnostic performance metrics and Cohen’s kappa (κ) were calculated. Results: Stroke detection showed a sensitivity of 100% (95% confidence interval [CI], 96.6–100%), specificity of 83.6% (95% CI, 75.6–89.4%), accuracy of 91.8%, and κ = 0.836. Accuracy was 93.6% for anterior/posterior circulation (κ = 0.838) and 90.9% for vascular territory (κ = 0.799). Conditional MCA segment-associated territory accuracy was 71.6% (53/74; κ = 0.602); end-to-end accuracy was 66.3% (53/80). Exploratory analysis of the 18 false-positive cases identified four overlapping error categories: artifactual and physiologic signal misinterpretation (72.2%), chronic lesion misclassification (27.8%), terminological inconsistency or unsupported report statements (77.8%), and etiological scenarios or unhedged definitive diagnoses (38.9%). Conclusions: GPT-5 achieved high sensitivity and a specificity of 83.6% for stroke classification on preselected representative DWI images using a vascular territory-based diagnostic framework. However, the 16.4% false-positive rate indicates that positive findings require radiologist confirmation. These findings support further evaluation of GPT-5 as a potential clinician-assisted image-interpretation tool but do not establish a clinical decision-support role.