Deep learning-based segmentation of the ischemic core and severe hypodensity on noncontrast head CT compared with CT perfusion and diffusion-weighted imaging.
Authors
Affiliations (9)
Affiliations (9)
- Department of Radiology, Stanford School of Medicine, Stanford, CA, USA. [email protected].
- Department of Radiology, Stanford School of Medicine, Stanford, CA, USA.
- Department of Neurology, Stanford School of Medicine, Stanford, CA, USA.
- Department of Neurology, Montpellier University Hospital, Montpellier, France.
- Department of Neurology, University Hospitals Leuven, Leuven, Belgium.
- Department of Electrical Engineering (ESAT), KU Leuven, Leuven, Belgium.
- Department of Medical Imaging, Sunnybrook Health Sciences Centre, University of Toronto, Toronto, ON, Canada.
- Department of Clinical Neurosciences, Lausanne University Medical Center, Lausanne, Switzerland.
- Department of Neuroradiology, MD Anderson, Houston, TX, USA.
Abstract
To compare the spatial overlap, imaging marker agreement, and outcome associations of acute ischemic regions of interest (ROIs) using deep-learning NCCT (DLNCCT) versus manual NCCT, CTP, and DWI-based segmentations. We trained a deep learning model to segment ischemic ROIs on admission NCCT (DLNCCT). DLNCCT was compared with manual NCCT, CTP (rCBF < 30%/38%), and DWI using the Dice similarity coefficient (mean ± SD), Bland-Altman analysis (mean difference [95% CI]), and concordance correlation coefficient (CCC). For each ROI, we compared NCCT imaging biomarkers: average lesion density (HU); modified net water uptake (mNWU in %); total volume (mL); and hypodense (≤ 26 HU) volume (mL). DLNCCT biomarkers were associated with reversed 90-day ordinal modified Rankin Scale (genOR [95% CI]). We used 218/762 train/test cases. Dice was 0.30 ± 0.30 between DLNCCT and manual segmentation, 0.22 ± 0.25 between DLNCCT-DWI, and 0.15 ± 0.22/0.21 ± 0.24 between CTP (rCBF < 30%/< 38%)-DWI. DLNCCT-DWI biomarker agreement was -1 HU (95% CI: -7; 6) for average density (CCC: 0.71), 4.9% (95% CI: -7.0; 16.8) for mNWU (CCC: 0.35), -16 mL (95% CI: -108; 76) for total volume (CCC: 0.57), and -4 mL (95% CI: -31; 23) for hypodense lesion volume (CCC: 0.75). Average lesion density (1.11 [95% CI: 1.06; 1.16]) and hypodense volume (0.85 [95% CI: 0.72; 0.99]) were associated with 90-day mRS. Spatial and biomarker agreement between DLNCCT and DWI was suboptimal but comparable to CTP-DWI. Nevertheless, DLNCCT-derived average lesion density and hypodense volume were associated with functional outcomes. Question Deep-learning models segment ischemic tissue on NCCT, but their spatial overlap and accuracy of imaging biomarker assessment relative to DWI and CTP core remain unclear. Findings Deep-learning NCCT segmentation showed good agreement with DWI in measuring hypodense (≤ 26 HU) volumes (- 4 mL [95% CI: -31; 23]; CCC 0.75). Clinical relevance Despite limited spatial overlap, NCCT-derived deep learning biomarkers showed agreement comparable to DWI, with improved performance in patients with poor ASPECTS and collateral scores. Hypodense volume and average lesion density were associated with functional outcome.