An adversarially consensus model of augmented unlabeled data for cardiac image segmentation (CAU⁺)

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• High-quality medical images are crucial for intelligent medical analysis.
• Acquiring professionally annotated medical image datasets is expensive and time-consuming.
• Semi-supervised learning offers a potential solution to reduce annotation dependency.

Purpose of the Study:

• To propose CAU+, a novel semi-supervised method for cardiac image segmentation.
• To leverage augmented unlabeled data to improve segmentation performance.
• To address the challenges of data scarcity and annotation costs in medical imaging.

Main Methods:

• CAU+ employs a consensus model with a segmentation network (teacher-student model) and a discriminator network.
• It utilizes CTAugment for strong and weak augmentation of unlabeled data, feeding them to student and teacher models, respectively.
• A hybrid loss function combines supervised and unsupervised losses, enhanced by adversarial learning using a discriminator-generated confidence map.

Main Results:

• CAU+ demonstrated significant improvements on the Automated Cardiac Diagnosis Challenge (ACDC) dataset.
• The method achieved up to an 18.01% increase in Dice Coefficient (DSC) and a 16.72% increase in Jaccard Coefficient (JC).
• Relative Absolute Volume Difference (RAVD) improved by up to 0.8, with Average Surface Distance (ASD) and 95% Hausdorff Distance (HD95) reduced by over 50% compared to state-of-the-art methods.

Conclusions:

• CAU+ is an effective and generalizable semi-supervised method for cardiac image segmentation.
• The proposed approach significantly enhances segmentation accuracy and reduces reliance on labeled data.
• CAU+ offers a promising solution for cost-effective and efficient medical image analysis.