Abstract WP103: Comparison Of Acute Infarct Lesions Between Non-contrast CT, DWI And FLAIR Using Back-to-back Imaging

Introduction: Size and location of the acute infarct is a major determinant of stroke outcome and eligibility for therapy. Recently, there have been efforts to train deep learning networks to detect lesions on Non-Contrast CT (NCCT) using concurrent DWI imaging as the gold standard. However, little is known about the radiological correspondence between concurrent NCCT and DWI lesion sizes. We performed an exploratory analysis comparing the stroke lesion volume on acute NCCT to that on DWI and FLAIR images performed shortly after. Methods: Population: DEFUSE 3 trial patients scanned 6-16h after last known well with DWI and NCCT <1h apart. NCCT was segmented using an algorithm detecting more than 5% depression compared to the contralateral homologous region. The DWIs and FLAIR were co-registered to NCCT and manually segmented in ITK-Snap. Results: Thirteen patients fulfilled inclusion criteria. DWI volumes were (median, [IQR]) 52 mL [39 - 102], FLAIR volumes 35 mL [18 - 66] and NCCT lesion volumes 14 mL [8 - 22]. NCCT lesions were significantly smaller than both DWI and FLAIR, p < 0.001 and p < 0.048. The Dice coefficient was 0.2 [0.1 - 0.3] for FLAIR vs NCCT and 0.3 [0.1 - 0.5] for DWI vs NCCT. Conclusion: NCCT lesion volumes were consistently smaller than both DWI and FLAIR lesions although the volumetric agreement with FLAIR lesions was better. Both FLAIR and NCCT are sensitive to influx of parenchymal water and as such may have more similar lesions than DWI imaging which is sensitive to early occurring intraparenchymal displacements. The findings suggest that the training of deep learning networks to detect early NCCT hypodensity should focus on FLAIR or expert outlines drawn directly on the NCCT rather than DWI as the gold standard.