Accuracy of Automated Deep Learning versus Expert Clinicians for Diagnosis of Acute Lacunar Stroke on CT Perfusion

BACKGROUND: Accurate diagnosis of lacunar stroke in the acute setting is challenging and often depends on MRI or clinical suspicion. CT perfusion (CTP) has only modest diagnostic accuracy and substantial interobserver variability yet is widely available. Artificial intelligence (AI) based tools may therefore be a valuable decision-support aid to improve the reliability of CTP interpretation. MATERIALS AND METHODS: We conducted a retrospective diagnostic accuracy study comparing a novel automated deep learning model with expert stroke neurologists for the diagnosis of acute lacunar stroke on CTP. Adults presenting to two comprehensive stroke centers in New South Wales, Australia, with a clinical syndrome suspicious for lacunar stroke who underwent standard-of-care CTP and diffusion-weighted MRI (DWI) were included. The reference standard was the presence of a clinically relevant lacunar infarct on DWI. A predetermined 90% of cases were used to train and internally test a convolutional neural network to predict the presence of a DWI lesion, while the remaining 10% formed a held out validation dataset. These held-out cases were interpreted by stroke neurologists, blinded to DWI, and by the trained model. Accuracy metrics, including area under the receiver operating characteristic curve (AUC), sensitivity, and specificity, were calculated for individual neurologists, neurologist consensus, and the AI model. RESULTS: Among 485 eligible patients, 436 (239 lacunar strokes and 197 mimics) were used for model development. The remaining 49 patients (27 lacunar strokes, 22 mimics) comprised the held out validation dataset. Overall clinician accuracy was poor (mean AUC 0.58) with low interobserver agreement (Fleiss κ = 0.22). The best-performing AI model, using cerebral blood flow maps alone, outperformed neurologist consensus (AUC 0.82, p = 0.003). CONCLUSION: In patients with suspected acute lacunar stroke, a deep learning model applied to routinely acquired CTP demonstrated diagnostic accuracy comparable to or better than expert stroke neurologists. These findings support further prospective and external validation and exploration of the model as a decision support tool within acute stroke workflows.