Multi-grained contrastive representation learning for label-efficient lesion segmentation and onset time...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate ischemic lesion segmentation and time since stroke (TSS) classification are vital for acute ischemic stroke (AIS) treatment, particularly tPA thrombolysis.
• Current deep learning models struggle with generalization due to limited paired MRI data, often overfitting to irrelevant features.

Purpose of the Study:

• To develop a deep learning framework that leverages readily available unpaired unlabeled MRI data to improve AIS analysis.
• To enhance the accuracy of ischemic lesion segmentation and TSS classification for unwitnessed AIS patients.

Main Methods:

• Proposed a multi-grained contrastive learning (MGCL) framework to learn representations from both coarse and fine-grained feature levels from unpaired data.
• Developed a novel multi-task framework for simultaneous lesion segmentation and TSS classification using limited labeled data.
• Introduced a multi-modal region-related feature fusion module to synergize features from different MRI modalities.

Main Results:

• The proposed framework demonstrated superior performance in ischemic lesion segmentation and TSS classification on a large-scale, multi-center MRI dataset.
• The MGCL framework effectively learned task-related prior representations, improving model generalization.
• The multi-task and feature fusion modules enhanced the accuracy of stroke evaluation and treatment decision-making.

Conclusions:

• The novel framework effectively utilizes unpaired unlabeled data to overcome limitations of supervised learning in AIS analysis.
• This approach shows significant promise for improving stroke evaluation and guiding treatment decisions in clinical practice.
• The integration of multi-grained learning and multi-modal feature fusion offers a robust solution for complex neuroimaging tasks.