Acute and sub-acute stroke lesion segmentation from multimodal MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Acute stroke lesion segmentation is crucial for timely treatment decisions.
• Magnetic Resonance Imaging (MRI) is the gold standard but time-consuming.
• Automated segmentation aids in assessing lesion location and volume, improving risk evaluation.

Purpose of the Study:

• To develop a deep learning methodology for segmenting acute and sub-acute stroke lesions using multimodal MRI.
• To address challenges like class imbalance and reduce post-processing needs.
• To evaluate the method's performance against state-of-the-art techniques.

Main Methods:

• A deep learning approach using multimodal MRI data.
• Pre-processing data leveraging brain hemisphere symmetry.
• Employing balanced patch sampling and a dynamically weighted loss function to handle class imbalance.
• Utilizing a U-Net based CNN architecture with overlapping patches for prediction.

Main Results:

• The method was evaluated on the ISLES 2015 challenge datasets for sub-acute stroke lesion segmentation (SISS) and acute stroke penumbra estimation (SPES).
• Achieved top rankings in online evaluations for both SISS (DSC=0.59 ± 0.31) and SPES (DSC=0.84 ± 0.10) tasks.
• Demonstrated superior performance compared to other methods, evidenced by a lower Hausdorff distance.

Conclusions:

• The proposed method effectively segments acute stroke lesions by integrating anatomical and pathophysiological information.
• The single training procedure demonstrates generalizability across different tasks and MRI modalities without hyper-parameter tuning.
• A public release of the method promotes scientific reproducibility.