Logo image
Back
Abstract DP364: Association of deep learning segmented ischemic core hypodensity on non-contrast CT with endovascular treatment benefit
Abstract   Peer reviewed

Abstract DP364: Association of deep learning segmented ischemic core hypodensity on non-contrast CT with endovascular treatment benefit

Henk van Voorst, Vignan Yogendrakumar, Hannah Johns, leonid churilov, Ameer Hassan, Michael Abraham, Santiago Ortega-Gutierrez, Muhammad Hussain, Michael Chen, Scott Kasner, …
Stroke (1970), Vol.57(Suppl_1), DP364
02/2026
DOI: 10.1161/str.57.suppl_1.DP364

View Online

Abstract

Background: Endovascular treatment (EVT) is effective for patients with large ischemic cores caused by a large vessel occlusion. The size and severity of hypodense tissue on non-contrast CT (NCCT) in the ischemic core predict poor outcomes despite successful EVT. Automated ASPECTS tools can estimate the size and severity of hypodensity on NCCT, but these estimates do not accurately represent the ischemic core. Deep learning (DL) based segmentation of hypodensity on NCCT may offer improved precision. We compared the associations between (1) ischemic core volume on CTP, (2) manually segmented, or (3) DL segmented hypodensity volumes on NCCT and outcomes after EVT or medical management (MM) in patients with a large ischemic core. Methods: We performed a post-hoc analysis of SELECT2 trial patients. The volume of total hypodense (all segmented voxels) and severely hypodense (segmented voxels with ≤26 HU) tissue was measured on baseline NCCT using an externally validated DL model or manual segmentations (Figure 1). The ischemic core volume was determined using CTP (rCBF<30%). We assessed agreement between DL and manually delineated volumes with the concordance correlation coefficients (CCC). Zou's Modified Poisson regression adjusted for age and baseline NIHSS was used to estimate the association of volume measures with 90-day independent ambulation (modified Rankin Scale ≤3), stratified by MM and EVT arms. Interaction p-values assessed whether volume measures modified EVT benefit. Results: 316 patients were included (154 MM, 162 EVT). Agreement between manual and DL segmented volumes was (CCC: 0.73[95%CI:0.67-0.78]) for total hypodense volume and (CCC: 0.90[95%CI:0.88-0.92]) for severely hypodense volume (Figure 2A-2B). Although manual and DL total hypodense volume on NCCT and ischemic core volume on CTP or MRI were associated with lower proportions of independent ambulation after EVT, we observed no association with EVT benefit (rows 1-3 Table 1). Severely hypodense volumes determined with either DL or manual segmentations were associated with a lower chance of independent ambulation after EVT and were significantly associated with reduced EVT benefit (rows 4-5 Table 1, Figure 2C-D). Conclusion: Severely hypodense volumes on NCCT are comparable between automated deep learning and manual segmentations. Our findings suggest that automated deep learning quantification of severe hypodensity on NCCT can help to identify patients less likely to benefit from EVT.
 Computed Tomography Deep learning Infarct size Endovascular Therapy Ischemic stroke

Details

Metrics

1 Record Views
Logo image