Exploiting Cross-modal Collaboration and Discrepancy for Semi-supervised Ischemic Stroke Lesion Segmentation from...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate ischemic stroke lesion segmentation is crucial for effective reperfusion therapy and understanding stroke causes.
• Multi-sequence MRI, including diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) maps, offers complementary information for improved segmentation.
• Current deep learning methods demand extensive annotated multi-modal MRI datasets, which are often impractical to acquire.

Purpose of the Study:

• To explore semi-supervised learning for ischemic stroke lesion segmentation from multi-sequence MRI, utilizing unlabeled data to enhance performance with limited annotations.
• To propose a novel framework that exploits cross-modality collaboration and discrepancy for efficient utilization of unlabeled data.

Main Methods:

• A semi-supervised framework for stroke lesion segmentation using multi-sequence MRI data.
• Implementation of a cross-modal bidirectional copy-paste strategy for inter-modality information exchange.
• Application of a cross-modal discrepancy-informed correction strategy to leverage limited labeled and abundant unlabeled data.

Main Results:

• The proposed method demonstrated efficient utilization of unlabeled data.
• Achieved 12.32% Dice Similarity Coefficient (DSC) improvement compared to a supervised baseline with only 10% annotations.
• Outperformed existing semi-supervised segmentation methods on the ISLES 22 dataset.

Conclusions:

• The developed semi-supervised framework effectively enhances ischemic stroke lesion segmentation accuracy by utilizing unlabeled multi-sequence MRI data.
• The cross-modality collaboration and discrepancy strategies enable efficient learning from limited annotations.
• This approach shows significant potential for improving clinical decision-making in stroke management.