Related Experiment Video
Updated: Oct 6, 2025
08:13
In Vivo Quantitative Assessment of Myocardial Structure, Function, Perfusion and Viability Using Cardiac Micro-computed Tomography
Published on: February 16, 2016
19.7K
U-net Models Based on Computed Tomography Perfusion Predict Tissue Outcome in Patients with Different Reperfusion
Yaode He1, Zhongyu Luo1, Ying Zhou1
1Department of Neurology, School of Medicine, the Second Affiliated Hospital of Zhejiang University, 88# Jiefang Road, Hangzhou, 310009, China.
Translational Stroke Research
|January 19, 2022
Summary
A novel deep learning U-net model accurately identifies brain tissue at risk and infarct core in acute large vessel occlusion (LVO) stroke patients using computed tomography perfusion (CTP) imaging, outperforming traditional methods.
Area of Science:
- Neuroimaging
- Artificial Intelligence in Medicine
- Stroke Imaging
Background:
- Accurate assessment of cerebral perfusion is crucial for guiding treatment decisions in acute large vessel occlusion (LVO) stroke.
- Identifying the ischemic core and tissue at risk is essential for predicting outcomes and selecting appropriate therapies.
Purpose of the Study:
- To develop and evaluate a deep learning U-net model for more accurate identification of ischemic core and tissue at risk using computed tomography perfusion (CTP) imaging.
- To compare the performance of the U-net model against traditional fixed-thresholding methods in patients with varying reperfusion patterns.
Main Methods:
- Development of two U-net deep learning models using baseline CTP images from 110 acute ischemic stroke patients.
- One model was trained on patients with major reperfusion (≥80%) to identify infarct core.
- The second model was trained on patients with minimal reperfusion (≤20%) to identify tissue at risk.
Main Results:
- The U-net model demonstrated superior performance in identifying infarct core (Dice score coefficient [DSC] 0.61, area under the curve [AUC] 0.92) compared to fixed-thresholding methods (DSC 0.52) in the major reperfusion group.
- In the minimal reperfusion group, the U-net model showed better performance for tissue at risk identification (DSC 0.67, AUC 0.93) versus fixed-thresholding (DSC 0.51).
- Excellent volumetric consistency was achieved, with intraclass correlation coefficients of 0.951 and 0.746 for major and minimal reperfusion groups, respectively.
Conclusions:
- CTP-based U-net models significantly outperform fixed-thresholding methods in identifying infarct core and tissue at risk in anterior LVO patients.
- These deep learning models offer a more accurate and individualized prediction of final infarct volume based on baseline CTP.
- The developed U-net models have the potential to improve treatment selection and patient outcomes in acute stroke management.
