Related Experiment Video
Updated: Jun 24, 2025
07:34
Author Spotlight: Establishing a Reliable Distal MCA Occlusion Model in Mice for Stroke Research
Published on: December 15, 2023
1.9K
A Deep Learning Approach to Predict Recanalization First-Pass Effect following Mechanical Thrombectomy in Patients
Haoyue Zhang1,2, Jennifer S Polson1,2, Zichen Wang1,2
1From the Computational Diagnostics Lab (H.Z., J.S.P., Z.W., W.F.S., C.W.A.), University of California, Los Angeles, California.
AJNR. American Journal of Neuroradiology
|June 13, 2024
Summary
Deep learning models can now predict the first-pass effect in stroke patients undergoing thrombectomy using CT and MR imaging. This automated approach eliminates the need for manual segmentation, improving prediction accuracy for better patient outcomes.
Area of Science:
- Neurology
- Medical Imaging
- Artificial Intelligence
Background:
- Successful recanalization in one attempt (first-pass effect) after endovascular thrombectomy for large-vessel occlusion stroke correlates with better long-term outcomes.
- Pretreatment imaging may hold predictive information for the first-pass effect.
- Existing machine learning models show promise but require manual segmentation.
Purpose of the Study:
- To develop fully automated deep learning methods for predicting the first-pass effect.
- To utilize pretreatment CT and MR imaging for prediction.
- To eliminate the need for manual segmentation in predicting recanalization outcomes.
Main Methods:
- A cohort of 326 patients undergoing endovascular thrombectomy was used.
- A hybrid transformer model with nonlocal and cross-attention modules was designed.
- The model was developed to predict the first-pass effect from MR imaging and CT series.
Main Results:
- The model achieved high cross-validation ROC-AUC: 0.8506 for MR imaging and 0.8719 for CT.
- On prospective test sets, ROC-AUC was 0.7967 for MR imaging and 0.8051 for CT.
- This represents the first classification of the first-pass effect from MR imaging alone and the first automated method for CT.
Conclusions:
- Nonperfusion MR imaging and CT signals can predict a successful first-pass effect.
- Deep learning methods enable prediction without time-intensive manual segmentation.
- These automated methods can aid in predicting outcomes for endovascular thrombectomy patients.
