Related Experiment Video
Updated: Jan 1, 2026
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
49.2K
A multi-path 2.5 dimensional convolutional neural network system for segmenting stroke lesions in brain MRI images
Yunzhe Xue1, Fadi G Farhat1, Olga Boukrina2
1Department of Computer Science, New Jersey Institute of Technology, Newark, NJ 07102, USA.
Neuroimage. Clinical
|December 23, 2019
Summary
This study introduces an automated system for identifying brain lesions in stroke survivors using magnetic resonance imaging (MRI). The novel multi-modal convolutional neural network significantly improves stroke lesion segmentation accuracy compared to existing methods.
Area of Science:
- Neuroimaging and Artificial Intelligence
- Medical Image Analysis
- Stroke Research
Background:
- Accurate identification of brain lesions in stroke survivors is crucial for diagnosis, treatment planning, and studying brain-behavior relationships.
- Manual lesion segmentation is time-consuming and labor-intensive, necessitating automated solutions.
- Existing automated methods struggle with accuracy, especially for smaller or varied lesion types.
Purpose of the Study:
- To develop and validate a multi-modal convolutional neural network (CNN) system for automated stroke lesion segmentation from MRI scans.
- To compare the proposed system's performance against previous state-of-the-art methods using cross-study validation.
Main Methods:
- A multi-modal, multi-path CNN system employing nine end-to-end UNets processing 2D MRI slices across three planes and normalizations.
- Outputs from the UNets are concatenated into a 3D volume and processed by a 3D CNN for final lesion mask generation.
- The system was trained and tested on diverse datasets (MCW, KF, ATLAS) including subacute/chronic and hemorrhagic/ischemic lesions.
Main Results:
- The proposed system achieved a mean Dice coefficient of 0.54 when trained on KF/MCW and tested on ATLAS, outperforming the best previous model (UNet at 0.47).
- Cross-study validation with combined datasets showed reliably higher accuracy than previous methods.
- The system demonstrated improved performance in identifying smaller lesions, a common challenge for existing techniques.
Conclusions:
- The developed multi-modal CNN system represents a significant advancement in automated stroke lesion segmentation.
- It offers improved accuracy and robustness across different datasets and lesion types, approaching inter-rater accuracy levels of human experts.
- This automated approach facilitates clinical diagnosis, treatment planning, and neuroscientific research in stroke.
