St-RegSeg: an unsupervised registration-based framework for multimodal magnetic resonance imaging stroke lesion...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Ischemic stroke is a leading cause of death and disability globally.
• Accurate assessment of stroke lesion size and location is critical for treatment, particularly for urgent vascular reconstruction surgery.
• Deep learning (DL) for multimodal MRI segmentation faces challenges integrating information and handling semantic drift.

Purpose of the Study:

• To propose the stroke unsupervised registration and segmentation (St-RegSeg) framework.
• To address limitations in current DL models for multimodal stroke MRI segmentation.
• To improve the accuracy and efficiency of registering and segmenting stroke lesions.

Main Methods:

• The St-RegSeg framework integrates unsupervised registration (ConvNXMorph) and segmentation (nnUNet-v2) models.
• It processes multimodal MRI images for both registration and segmentation tasks.
• The framework was evaluated on the ISLES'22 dataset from three centers.

Main Results:

• St-RegSeg significantly improved registration Dice Similarity Coefficient (DSC) by 25.31% and reduced Mean Squared Error (MSE) by 17.36% compared to ANTs+nnUNet-v2.
• Segmentation DSC improved by 0.84%, and overall inference speed increased by 40.91 times.
• Compared to TransMorph+nnUNet-v2, St-RegSeg showed improvements in DSC (3.68%), MSE (8.91%), NCC (8.49%), MI (6.18%), and segmentation DSC (0.5%), with a 2.13x speed increase.

Conclusions:

• The St-RegSeg framework offers a highly effective solution for multimodal MRI registration and segmentation in ischemic stroke.
• It demonstrates superior performance metrics and computational efficiency compared to existing methods.
• The open-sourced framework presents a promising tool for clinical applications in stroke management.