LCC-Net: A Lightweight Cross-Consistency Network for Semisupervised Cardiac MR Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Semantic segmentation is vital for cardiac magnetic resonance (MR) image analysis.
• Supervised deep learning methods require extensive pixel-wise annotations, which are scarce in clinical settings.
• Current high-performing models often exhibit high computational complexity.

Purpose of the Study:

• To develop an efficient cardiac image segmentation method for low-data scenarios.
• To introduce a lightweight cross-consistency network (LCC-Net) addressing annotation and computation constraints.
• To leverage abundant unlabeled data for improved segmentation performance.

Main Methods:

• Implemented a lightweight module replacing standard convolutions to mitigate overfitting on small datasets.
• Introduced extreme consistency learning to enforce equivariant constraints on perturbed image versions.
• Utilized cutting and mixing of training images as an extreme perturbation for robust representation learning.

Main Results:

• LCC-Net demonstrates promising performance with high annotation- and computation-efficiency.
• Achieved a 14.4% gain in mean Dice score over the baseline U-Net with only two annotated subjects.
• The model effectively leverages unlabeled data through extreme consistency learning.

Conclusions:

• LCC-Net offers an effective solution for cardiac image segmentation with limited labeled data.
• The proposed method significantly reduces computational complexity and annotation requirements.
• Extreme consistency learning enhances robustness and performance in data-scarce environments.