Deep learning-based automated lesion segmentation on mouse stroke magnetic resonance images
Jeehye An1,2, Leo Wendt3, Georg Wiese3
1Department of Experimental Neurology and Center for Stroke Research, Charité-Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Berlin, Germany.
Scientific Reports
|August 16, 2023
Summary
We developed an automated deep learning method for segmenting ischemic stroke lesions in mouse MRI scans. This approach significantly reduces manual labor and variability, achieving high accuracy and standardization in lesion detection.
Area of Science:
- Neuroimaging
- Biomedical Engineering
- Artificial Intelligence
Background:
- Manual segmentation of ischemic stroke lesions in mouse MRI is labor-intensive and prone to variability.
- Accurate lesion quantification is crucial for preclinical stroke research.
Purpose of the Study:
- To develop and validate a fully automated deep learning method for ischemic stroke lesion segmentation in mouse T2-weighted MRI.
- To improve efficiency and reduce inter- and intra-rater variability in lesion analysis.
Main Methods:
- An end-to-end deep learning model was trained on 293 mouse MRI scans for automated lesion segmentation.
- The model processed raw MRI data with minimal preprocessing.
- Performance was evaluated using Dice coefficients and lesion volume accuracy against manual segmentation on an independent dataset.
Main Results:
- The automated method produced smooth, compact, and realistic segmentation masks.
- Achieved high agreement with manual segmentation, surpassing inter-rater reliability reported in prior studies.
- Demonstrated robust performance on an independent dataset with varying imaging characteristics.
Conclusions:
- Fully automated deep learning offers a standardized and efficient alternative to manual lesion segmentation in mouse stroke models.
- This method has the potential to reduce bias and improve reproducibility across research studies and institutions.
- Enables more reliable quantification of ischemic stroke lesions in preclinical research.
