Deep learning-based detection and segmentation of diffusion abnormalities in acute ischemic stroke

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neurology

Background:

• Clinical and research communities require efficient tools for detecting and segmenting diffusion abnormalities in acute strokes.
• Current methods may lack the speed, accuracy, or accessibility needed for large-scale studies.

Purpose of the Study:

• To develop and validate an accessible deep learning tool for accurate detection and segmentation of acute stroke lesions.
• To provide a robust and efficient solution for analyzing diffusion-weighted MRIs in stroke patients.

Main Methods:

• Developed a deep learning tool trained on 2,348 clinical diffusion-weighted MRIs of acute and sub-acute ischemic strokes.
• Validated the tool's generalization on an external dataset of 280 MRIs (STIR dataset).
• Compared performance against generic networks and DeepMedic.

Main Results:

• The tool demonstrated superior performance over generic networks and DeepMedic, especially for small lesions.
• Achieved a lower false positive rate, balanced precision and sensitivity, and robustness to data perturbations.
• Automated lesion quantification showed near-perfect agreement with human measurements, rivaling inter-evaluator agreement.

Conclusions:

• The developed tool is fast, public, and requires minimal computational resources, making it accessible to non-experts.
• It meets the criteria for performing large-scale, reliable, and reproducible clinical and translational research in stroke.
• Enables efficient and accurate analysis of diffusion abnormalities in acute ischemic stroke.