Clinically Scalable Deep Learning for Stroke Multicenter Validation of an Ultra-Efficient NCCT-Based Diagnostic Framework

We propose an ultra-efficient multitask deep learning framework for automated cerebral ischemia detection and lesion segmentation from non-contrast computed tomography (NCCT) scans. Evaluated on 1,200 multicenter NCCT scans using fivefold cross-validation and an external test set, the model achieved 97.2% accuracy, 97.8% sensitivity, and an area under the curve (AUC) of 0.984 for classification, along with a Dice coefficient of 0.842 and intersection-over-union (IoU) of 0.730 for segmentation. Notably, the model requires only ~1.8 million parameters and achieves an average inference time of 0.92 seconds per scan, making it computationally lightweight. By integrating binary classification and voxel-level segmentation into a single interpretable framework, this approach addresses critical clinical needs for rapid, accurate, and scalable stroke assessment. The strong performance and generalizability across multicenter data suggest its potential for real-time integration into clinical workflows, warranting future prospective validation.