Abstract TP64: Ischemic Stroke Lesion Identification in Non-Contrast CT Using Deep Learning

Background: Automatic measurement of the acute stroke lesion volume on DWI and CT-CBF has been used in recent late window trials. Despite non-contrast CT (NCCT) being the most widely used imaging modality in the acute stroke setting, quantification of acute stroke volumes on NCCT has not been employed in trials because of the difficulty outlining territory with very mild Hounsfield unit depression. Deep learning algorithms have been effective at solving many image processing tasks and may outperform human readers given enough training data. The goal of this study was to train and test a deep learning model on NCCT scans with synthetic stroke lesions and to determine the optimal model design. Methods: Training: 20 NCCT scans without acute stroke were combined with 20 DWI lesions using co-registration producing 400 non-contrast scans with lesions. The region of the NCCT that coincided with the DWI lesion was depressed by 2 Hounsfield units to simulate an acute infarct. An independent validation dataset of 100 cases was created in the same way. Two models were used: a standard “Unet” model and a symmetry-aware Unet model. The models were compared in terms of segmentation accuracy in the independent validation dataset. Results: Both the symmetry aware U-net and the standard U-net detected some part of the true lesion in 100% of the cases. The symmetry aware U-net was more sensitive, median [iqr], (45% [27-68] vs 17% [6-54], p<0.00001) but slightly less specific (98% [93-98] vs 99% [94-99], p<0.0008) than the standard U-net. Conclusion: The symmetry aware U-net shows great promise in detection of acute strokes on NCCT; lesions with Hounsfield unit depressions that are barely visible to the eye can be automatically segmented by this model. Additional training data and architectural enhancements are likely to improve the current spatial sensitivity to above 45%.