A deep learning algorithm achieved 72% accuracy in detecting ischaemic stroke lesions on brain CT scans, according to a study published in Stroke and Vascular Neurology.

The study used data from the Third International Stroke Trial, analyzing 5,772 CT scans from 2,347 patients (median age 82 years) who were imaged within 6 hours of stroke onset. Of these scans, 54% showed positive findings for ischaemic lesions according to expert review.

The researchers developed a convolutional neural network (CNN)-based deep learning method to identify both the presence and location (left, right, or both sides) of ischaemic lesions, without requiring annotated images. Using a 70-15-15 split for training, validation, and testing, the model outperformed other tested architectures including Vision Transformer, Swin Transformer, and ResNet variants.

Results showed higher accuracy for follow-up scans (76%) compared to baseline scans (67%). The model demonstrated 80% specificity overall (83% follow-up, 79% baseline) and varying sensitivity (78% follow-up, 56% baseline). Performance improved with lesion size, achieving 80% accuracy for larger lesions. Multiple lesions showed increasing accuracy rates: 62% for single lesions, 87% for two lesions, and 100% for three or more lesions.

Regional accuracy varied, with best performance in the anterior cerebral artery (75%) and middle cerebral artery regions (68%). The system showed lower accuracy for brainstem (20%), lacunar (33%), and cerebellar lesions (33%). Background conditions impacted performance, with error rates of 31% for old stroke lesions and 32% for non-stroke lesions.

Key limitations included reduced visibility of culprit lesions on baseline CT scans and limited cases for certain lesion categories, affecting subgroup analyses. The study demonstrated the feasibility of developing AI systems using routine clinical scans rather than specially annotated research images, potentially enabling larger-scale algorithm development using real-world data.

Full disclosures can be found in the published study.