Semi-Supervised Medical Image Segmentation with Co-Distribution Alignment.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Medical image segmentation relies heavily on large labeled datasets, which are costly and time-consuming to create due to the need for expert annotation.
• Uneven class distribution in medical images often leads to poor performance on underrepresented categories, hindering diagnostic accuracy.
• Current semi-supervised methods struggle with the challenges of limited labeled data and imbalanced classes in medical image segmentation.

Purpose of the Study:

• To propose a novel semi-supervised approach, Co-Distribution Alignment (Co-DA), for medical image segmentation.
• To address the limitations of expensive annotation and class imbalance in medical image segmentation datasets.
• To enhance segmentation performance, particularly for minority classes, using limited labeled data.

Main Methods:

• Co-Distribution Alignment (Co-DA) aligns class-wise marginal predictions between two differently initialized models on unlabeled data.
• Pseudo-labels generated by one model are used to supervise the other, facilitating knowledge transfer.
• An over-expectation cross-entropy loss is introduced to filter noisy pseudo-labels from unlabeled pixels.

Main Results:

• Co-DA achieved superior performance compared to state-of-the-art methods on public 2D CaDIS, 3D LGE-MRI, and ACDC datasets.
• The method obtained an mIoU of 0.8515 on CaDIS using only 24% labeled data.
• Dice scores of 0.8824 on LGE-MRI and 0.8773 on ACDC were achieved with only 20% labeled data.

Conclusions:

• Co-Distribution Alignment (Co-DA) is an effective semi-supervised strategy for medical image segmentation, significantly reducing reliance on extensive annotations.
• The proposed method demonstrates robust performance even with limited labeled data and addresses the challenge of class imbalance.
• Co-DA offers a promising solution for improving the efficiency and accuracy of medical image segmentation in resource-constrained scenarios.