Abstract
Background And Aims
Background and Aims: In patients with ischemic stroke presenting with the hyperacute timeframe,
ischemia can be challenging to detect on non-contrast CT (NCCT) and consequently may be missed. We developed a deep-learned CT denoising model to improve contrast-to-noise ratio (CNR) in NCCT. We apply the model in hyperacute stroke lesions that were missed on the initial NCCT to quantitatively measure the improvement in lesion discrimination and visualization.
Methods
Methods: In this retrospective study, 18 patients (64±12 years, 8 men) with AIS that have 4-7 full-dose NCCT head scans within 5-day timeframe (83 scans) were selected from a stroke registry between 2016-2020. Each patientʼs scans were co-registered to generate a mean CT as a reference standard for training a deep learning model (rotation-reflection equivariant U-Net with group convolution (G-CNN) modification) to denoise new CT scans. The model denoising performance is tested in 307 patients. We further tested the model in 19 patients with hyperacute stroke missed in the initial NCCT but subsequently confirmed with MRI-DWI. CNR was measured between the stroke lesion and contralateral normal tissue, and the fold change in CNR after denoising was compared against the original CT.
Results
Results: After denoising, CNR of missed acute ischemic stroke versus contralateral normal tissue
improved by 2.2±0.7 folds. Deep learning denoised scans are visually clearer with preserved anatomy.
Background and Aims: In patients with ischemic stroke presenting with the hyperacute timeframe,
ischemia can be challenging to detect on non-contrast CT (NCCT) and consequently may be missed. We developed a deep-learned CT denoising model to improve contrast-to-noise ratio (CNR) in NCCT. We apply the model in hyperacute stroke lesions that were missed on the initial NCCT to quantitatively measure the improvement in lesion discrimination and visualization.
Methods
Methods: In this retrospective study, 18 patients (64±12 years, 8 men) with AIS that have 4-7 full-dose NCCT head scans within 5-day timeframe (83 scans) were selected from a stroke registry between 2016-2020. Each patientʼs scans were co-registered to generate a mean CT as a reference standard for training a deep learning model (rotation-reflection equivariant U-Net with group convolution (G-CNN) modification) to denoise new CT scans. The model denoising performance is tested in 307 patients. We further tested the model in 19 patients with hyperacute stroke missed in the initial NCCT but subsequently confirmed with MRI-DWI. CNR was measured between the stroke lesion and contralateral normal tissue, and the fold change in CNR after denoising was compared against the original CT.
Results
Results: After denoising, CNR of missed acute ischemic stroke versus contralateral normal tissue
improved by 2.2±0.7 folds. Deep learning denoised scans are visually clearer with preserved anatomy.
| Original language | English (US) |
|---|---|
| Pages | 1717 |
| DOIs | |
| State | Published - Sep 1 2021 |
| Event | European Stroke Organization Conference 2021: ESOC 2021 - Virtual this year Duration: Sep 1 2021 → Sep 3 2021 https://2021.eso-conference.org/ |
Conference
| Conference | European Stroke Organization Conference 2021 |
|---|---|
| Period | 9/1/21 → 9/3/21 |
| Internet address |