Automated deep U-Net model for ischemic stroke lesion segmentation in the sub-acute phase

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Manual segmentation of sub-acute ischemic stroke lesions in FLAIR MRI is labor-intensive and prone to variability.
• This inefficiency hinders timely clinical decision-making and patient management.

Purpose of the Study:

• To develop and validate an automated deep learning framework for accurate segmentation of sub-acute ischemic stroke lesions in FLAIR MRI.
• To improve the efficiency and reliability of stroke lesion analysis in clinical settings.

Main Methods:

• A novel multi-path residual U-Net architecture with 2.34 million parameters was proposed.
• Hyperparameters were optimized via 5-fold cross-validation, and N4 bias field correction was used.
• Rigorous patient-level data partitioning and bias-corrected bootstrap confidence intervals ensured robust validation.

Main Results:

• The model achieved a validation Dice Similarity Coefficient (DSC) of 0.85 ± 0.12 and demonstrated consistent performance on the test set (DSC: 0.89 ± 0.07).
• It showed statistically significant improvements over DRANet, 2D CNN, and Attention U-Net.
• Inference time was rapid at 45 ms per slice, highlighting computational efficiency.

Conclusions:

• The proposed deep learning framework offers robust and accurate automated segmentation of sub-acute ischemic stroke lesions.
• The model's performance and efficiency suggest potential for integration into clinical workflows.
• Further multi-site validation is recommended before widespread clinical implementation.