Swin Transformer Enhanced with OOD Detection for Robust and Reliable Diagnosis of Ischemic Stroke from CT Image

Diagnosing ischemic stroke from CT scan images presents significant challenges in achieving the speed and accuracy essential for clinical decision-making, where conventional CNN-based methods show limitations. This study addresses these gaps by developing an automated diagnostic system using a Swin Transformer model integrated with an Out-of-Distribution (OOD) detection mechanism to enhance diagnostic reliability. The model was trained and validated on a dataset of 583 brain CT images from 341 patients at a regional hospital in Makassar. This dataset, labeled by two expert radiologists (κ=0.94), was categorized into ischemic stroke (206), normal (228), and non-brain CT scans (149) as the OOD class. The Swin Transformer achieved an exceptional validation accuracy of 99.15% after 10 epochs, with a highly efficient total training time of approximately 24 minutes. The model’s superiority was further confirmed by high weighted averages for precision (0.99), recall (0.99), and F1-score (0.99). Critically, the OOD detection module demonstrated perfect performance, achieving 100% accuracy in identifying irrelevant images with a 0% false positive rate, thereby preventing erroneous diagnoses from non-brain scans. Robustness testing under varied lighting conditions also showed a 100% success rate. Real-time viability was confirmed through external validation using a live camera, yielding a rapid inference time of 0.3 seconds per image. This study concludes that the developed system offers a highly accurate, robust, and safe solution, proving its readiness for clinical implementation to support ischemic stroke diagnosis in Indonesia.