Related Experiment Video
Updated: Nov 7, 2025
09:59
A Magnetic Resonance Imaging Protocol for Stroke Onset Time Estimation in Permanent Cerebral Ischemia
Published on: September 16, 2017
14.3K
Intra-domain task-adaptive transfer learning to determine acute ischemic stroke onset time
Haoyue Zhang1, Jennifer S Polson1, Kambiz Nael2
1Computational Diagnostics Lab, University of California, Los Angeles, CA 90024, USA; Department of Bioengineering, University of California, Los Angeles, CA 90024, USA.
Summary
Deep learning models can estimate time since stroke onset (TSS) using MRI diffusion imaging for acute ischemic stroke (AIS) patients with unknown onset times. This approach aids thrombolysis decisions when TSS is uncertain.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Stroke Medicine
Background:
- Time since stroke onset (TSS) is critical for acute ischemic stroke (AIS) treatment decisions.
- TSS is unknown in up to 25% of AIS cases, often due to unwitnessed strokes.
- Current MRI assessment for unknown TSS has significant inter-reader variability, impacting treatment eligibility.
Purpose of the Study:
- To develop and validate deep learning (DL) models for classifying TSS using MRI diffusion series in AIS patients.
- To implement an intra-domain task-adaptive transfer learning method for improved DL model performance.
- To provide a more inclusive and accurate method for determining thrombolysis eligibility in patients with unknown TSS.
Main Methods:
- Developed 2D and 3D Convolutional Neural Network (CNN) architectures.
- Employed an intra-domain task-adaptive transfer learning approach, pre-training on stroke detection and fine-tuning on TSS thresholds.
- Applied models to a broad patient cohort, including diverse clinical, demographic, and imaging criteria.
Main Results:
- The top DL model achieved an ROC-AUC of 0.74, with 0.70 sensitivity and 0.81 specificity for classifying TSS < 4.5 hours.
- Pre-trained models outperformed models trained from scratch and previous published models on the same dataset.
- The DL pipeline demonstrated 75.78% overall accuracy in classifying TSS < 4.5 hours on a broad patient spectrum.
Conclusions:
- DL models leveraging MRI diffusion series can accurately classify time since stroke onset (TSS) for AIS patients with unknown onset.
- The proposed transfer learning method enhances classification performance, exceeding existing benchmarks.
- This approach offers a promising, inclusive tool to guide thrombolysis decisions in AIS patients with unknown TSS.
