Automatic post-stroke lesion segmentation on MR images using 3D residual convolutional neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Chronic stroke leads to irreversible brain tissue damage, requiring accurate volumetric assessment.
• T1-weighted Magnetic Resonance Imaging (MRI) is crucial for visualizing brain lesions.
• Accurate segmentation of these lesions is vital for understanding disease progression and treatment efficacy.

Purpose of the Study:

• To evaluate the feasibility and performance of deep residual neural networks for volumetric segmentation of brain lesions in chronic stroke patients.
• To assess the accuracy of 3D deep convolutional segmentation models using T1-weighted MRI scans.
• To compare the performance of deep learning models against manual lesion tracing.

Main Methods:

• Retrospective analysis of 239 T1-weighted MRI scans from chronic ischemic stroke patients.
• Application of 3D deep convolutional segmentation models with residual learning and a novel zoom-in&out strategy.
• Performance evaluation using Dice Similarity Coefficient (DSC), Average Symmetric Surface Distance (ASSD), and Hausdorff Distance (HD) with bootstrapping for confidence intervals.

Main Results:

• The deep learning models achieved an average DSC of 0.64 (median 0.78) on a test set of 31 scans.
• Average ASSD was 3.6 mm and average HD was 20.4 mm, indicating reasonable lesion boundary and volumetric accuracy.
• The models demonstrated effectiveness in segmenting irreversibly damaged brain tissue lesions.

Conclusions:

• Deep residual neural networks are feasible and perform effectively for volumetric segmentation of chronic ischemic stroke lesions on T1-weighted MRI.
• The applied 3D deep learning techniques show promise for automated and accurate lesion analysis in stroke research.
• This approach can aid in the quantitative assessment of brain tissue damage in chronic stroke patients.