Ischemic Lesion Segmentation using Ensemble of Multi-Scale Region Aligned CNN.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate ischemic stroke lesion segmentation is vital for effective treatment planning.
• Manual segmentation is time-consuming and subject to variability.
• Existing deep learning methods show promise but require enhancement.

Purpose of the Study:

• To propose an enhanced deep learning architecture for accurate ischemic stroke lesion segmentation from multimodal MRI.
• To improve upon existing automated segmentation methods by refining feature representation and model robustness.

Main Methods:

• Developed a novel deep architecture incorporating multi-level losses, multi-scale feature integration, and sub-network ensemble predictions.
• Introduced a custom dropout module for progressive feature refinement and attention mechanisms for selective feature weighting.
• Employed patch-based modeling and separate classification/segmentation branches to address data imbalance.

Main Results:

• The proposed architecture achieved a mean Dice coefficient of 0.775 on the ISLES 2015 SISS dataset.
• Demonstrated superior segmentation performance compared to existing models.
• The fully automated framework integrating classification and segmentation yielded improved results.

Conclusions:

• The enhanced deep learning architecture significantly improves automated ischemic stroke lesion segmentation accuracy.
• The proposed methods offer a robust and efficient solution for clinical applications.
• This work advances the state-of-the-art in medical image analysis for stroke diagnosis.