Automated ischemic stroke lesion detection on non-contrast brain CT: a large-scale clinical feasibility test AI stroke lesion detection on NCCT

This is totally fucking useless without the followup protocols that deliver recovery! Why are you stopping at the first step? Diagnosing something without a recovery plan is the height of stupidity!

Automated ischemic stroke lesion detection on non-contrast brain CT: a large-scale clinical feasibility test AI stroke lesion detection on NCCT

JoonNyung Heo¹Wi-Sun Ryu²Jong-Won Chung³Chi Kyung Kim⁴Joon-Tae Kim⁵Myungjae Lee²Dongmin Kim²Leonard Sunwoo⁶Johanna M. Ospel^7,8Nishita Singh⁹Hee-Joon Bae¹⁰Beom Joon Kim^10*

• ¹Department of Neurology, Yonsei University College of Medicine, Seoul, Republic of Korea
• ²Artificial Intelligence Research Center, JLK Inc., Seoul, Republic of Korea
• ³Department of Neurology, Samsung Medical Center, Sungkyunkwan University College of Medicine, Seoul, Republic of Korea
• ⁴Department of Neurology, Korea University Guro Hospital and Korea University College of Medicine, Seoul, Republic of Korea
• ⁵Department of Neurology, Chonnam National University Hospital, Chonnam National University Medical School, Gwangju, Republic of Korea
• ⁶Department of Radiology, Seoul National University Bundang Hospital, Seoul National University College of Medicine, Seongnam, Republic of Korea
• ⁷Department of Diagnostic Imaging, Foothills Medical Center, University of Calgary, Calgary, AB, Canada
• ⁸Department of Clinical Neurosciences, Foothills Medical Center, University of Calgary, Calgary, AB, Canada
• ⁹Department of Neurology, Faculty of Medicine, University of Manitoba, Winnipeg, MB, Canada
• ¹⁰Department of Neurology, Seoul National University Bundang Hospital, Seongnam, Republic of Korea

Background: Non-contrast CT (NCCT) is widely used imaging modality for acute stroke imaging but often fails to detect subtle early ischemic changes. Such underestimation can lead clinicians to overlook tissue-level information. This study aimed to develop and externally validate automated software for detecting ischemic lesions on NCCT and to assess its clinical feasibility in stroke patients undergoing endovascular thrombectomy.

Methods: In this retrospective, multicenter cohort study (May 2011–April 2024), a modified 3D U-Net model was trained using paired NCCT and diffusion-weighted imaging (DWI) data from 2,214 patients with acute ischemic stroke. External validation was performed in 458 subjects. Clinical feasibility was assessed in 603 endovascular thrombectomy-treated patients with complete recanalization. Model outputs were compared against expert-annotated DWI lesions for sensitivity, specificity, and volumetric correlation. Clinical endpoints included follow-up DWI lesion volumes, hemorrhagic transformation, and 3-month modified Rankin Scale outcomes.

Results: A total of 458 subjects were evaluated for external validation (mean age, 64 years ± 16; 265 men). The model achieved 75.3% sensitivity (95% CI, 70.9–79.9%) and 79.1% specificity (95% CI, 77.1–81.3%). In the feasibility cohort (n = 603; mean age, 69 years ± 13; 362 men), NCCT-derived lesion volumes correlated with follow-up DWI volumes (ρ = 0.60, p < 0.001). Lesions >50 mL were associated with reduced favorable outcomes (17.3% [26/150] vs. 54.2% [246/453], p < 0.001) and higher hemorrhagic transformation rates (66.0% [99/150] vs. 46.3% [210/453], p < 0.001). Radiomics features improved hemorrhagic transformation prediction beyond clinical variables alone (area under the receiver operating characteristic curve, 0.833 vs. 0.626; p = 0.003).

Conclusion: The automated NCCT-based lesion detection model demonstrated reliable diagnostic performance and provided clinically relevant prognostic information in endovascular thrombectomy-treated stroke patients.

Introduction

Non-contrast computed tomography (NCCT) is the most widely accessible imaging modality for acute stroke worldwide due to its accessibility and utility in rapidly ruling out hemorrhagic stroke (Kurz et al., 2016). However, hypodense changes indicative of acute ischemia can be subtle, leading to suboptimal sensitivity (van Horn et al., 2021). Although semiquantitative scores are routinely used to communicate the extent of ischemic changes, their inter-rater reliability varies considerably (Farzin et al., 2016). Likewise, manual segmentation of early ischemic changes on NCCT often yields low agreement (Christensen et al., 2023).

Despite these limitations, the initial NCCT scan contains useful tissue-level information such as extent and severity of ischemia, which has not been thoroughly utilized in clinical practice (Demeestere et al., 2025). Greater emphasis on detecting and quantifying ischemic changes could guide treatment decisions, especially for time-sensitive interventions such as intravenous thrombolysis and endovascular treatment (EVT).

We aimed to develop an automated software model to detect acute ischemic lesions on NCCT. The model’s training and validation used concomitant diffusion-weighted imaging (DWI) with expert ratings as a reference standard, given DWI’s high sensitivity for acute infarction. We further tested the software’s clinical feasibility in a separate cohort of patients with large vessel occlusion (LVO) undergoing EVT with complete recanalization, correlating NCCT-derived lesion volumes and radiomics features with subsequent DWI and clinical outcomes.

Methods

The study conformed to the Standards for Reporting of Diagnostic Accuracy Studies guidelines for diagnostic accuracy research (Cohen et al., 2016).

Acute ischemic lesion detection on NCCT: model development and validation

We retrospectively collected data from six stroke centers in South Korea between 2011 and 2015, including 2,398 ischemic stroke patients. Inclusion required adults with acute ischemic stroke who underwent NCCT and DWI within 3 h to minimize ischemic lesion evolution, while patients with poor image quality, structural brain abnormalities, or incomplete expert annotations were excluded (Figure 1A). Five expert neurologists each with over 10 years of clinical experience, manually annotated ischemic lesions visible on NCCT while referring to DWI / apparent diffusion coefficient images for confirmation. Experts’ annotation agreements were assessed using volumetric similarity indices and absolute volume difference, with ground truth defined by consensus of more than 2 experts. The labeled 2,398 ischemic stroke patients were randomly categorized into 2,214 cases for training / internal validation cohort and 184 cases for external validation cohort with 274 non-stroke individuals.