Deep learning for cerebral vascular occlusion segmentation: A novel ConvNeXtV2 and GRN-integrated U-Net framework for...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Cerebral vascular occlusion can cause stroke and permanent neurological damage.
• Accurate segmentation of brain lesions in Magnetic Resonance Imaging (MRI) is crucial for treatment.
• Challenges in MRI segmentation include low contrast, noise, and complex lesion structures.

Purpose of the Study:

• To develop a novel deep learning model for accurate cerebral vascular occlusion segmentation in MRI.
• To address limitations of existing methods, including computational cost and segmentation of small or irregular lesions.
• To enhance Computer-Aided Diagnosis (CAD) systems for stroke detection.

Main Methods:

• A novel U-Net architecture incorporating ConvNeXtV2 blocks and GRN-based Multi-Layer Perceptrons (MLP) was proposed.
• The model was applied to cerebral vascular occlusion segmentation, with preprocessing to remove small lesions (≤5 pixels).
• Experiments were conducted on the ISLES 2022 dataset.

Main Results:

• The proposed model achieved a high Intersection over Union (IoU) of 0.8015 and a Dice coefficient of 0.8894.
• Demonstrated significant improvement in segmentation accuracy, particularly in low-contrast regions.
• Showcased high computational efficiency suitable for clinical applications.

Conclusions:

• The novel U-Net architecture with ConvNeXtV2 and MLP offers superior performance for cerebral vascular occlusion segmentation.
• The model provides a promising, accurate, and efficient solution for clinical diagnosis and treatment planning.
• This work represents the first application of ConvNeXtV2 for this specific medical imaging task.