Reliability-Aware Semi-supervised Mutual Learning for Acute Ischemic Stroke Lesion Segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine
• Neurology

Background:

• Accurate lesion localization is crucial for acute ischemic stroke (AIS) treatment.
• Automatic stroke lesion segmentation is challenging due to limited annotated medical datasets.
• Semi-supervised learning (SSL) shows promise but is hampered by unreliable pseudo-labels.

Purpose of the Study:

• To propose a novel semi-supervised learning framework, Reliability-Aware Mutual Learning (RAML), for improved stroke lesion segmentation.
• To address the challenge of unreliable pseudo-labels in medical image segmentation using SSL.
• To enhance the accuracy and efficiency of automatic lesion detection in AIS.

Main Methods:

• Developed RAML, a framework with two subnetworks sharing an encoder and employing primary and auxiliary decoders.
• Introduced Uncertain Region Relearning (URR) to refine unreliable regions in labeled data using prediction uncertainty.
• Implemented Reliability-Aware Mutual Pseudo-Supervision (RMPS) for cross-supervision on unlabeled data using reliable pseudo-labels.
• Incorporated Feature Difference Learning (FDL) to promote prediction diversity between subnetworks.

Main Results:

• Demonstrated the effectiveness of RAML on two acute ischemic stroke datasets.
• Validated the framework's performance on the Left Atrium dataset, showcasing its versatility.
• RAML significantly improved semi-supervised segmentation tasks compared to existing methods.

Conclusions:

• RAML offers a robust solution for semi-supervised medical image segmentation, particularly for AIS.
• The proposed regularization techniques effectively handle unreliable pseudo-labels and enhance segmentation accuracy.
• This framework has the potential to improve clinical outcomes through more precise stroke lesion identification.