Semi-supervised cine cardiac MRI segmentation via joint registration and temporal attention perceiver

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Background:

• Accurate segmentation of cardiac structures in cine cardiac MRI is crucial for evaluating heart function.
• Deep learning models achieve high accuracy but require extensive, fully annotated datasets, which are labor-intensive to create.
• Semi-supervised methods reduce annotation burden but often yield suboptimal segmentation accuracy.

Purpose of the Study:

• To develop an improved semi-supervised segmentation method for cine cardiac MRI.
• The method aims to enhance segmentation accuracy using small training datasets and limited annotations.

Main Methods:

• A novel approach combining deformable registration, fully and weakly supervised segmentation, and a temporal attention perceiver (TAP) was developed.
• Deformable registration generated pseudo-labels for unlabeled data, which were used alongside labeled data for training.
• The TAP module refined features and enforced cross-instance feature alignment to improve registration accuracy.

Main Results:

• The proposed method achieved Dice-similarity-coefficients (DSC) of 0.910 ± 0.063 for RV, 0.894 ± 0.024 for Myo, and 0.934 ± 0.056 for LV on the ACDC dataset.
• Segmentation accuracies were comparable to a fully supervised U-Net model (Unet_UB) and outperformed the bidirectional copy-paste (BCP) method.
• Similar performance trends were observed on the Multi-Vendor & Multi-Disease (M&Ms) dataset, demonstrating robustness across different datasets.

Conclusions:

• The developed semi-supervised approach significantly outperforms existing methods in cine cardiac MRI segmentation.
• It achieves segmentation accuracies comparable to fully supervised models, even with limited training data and under-annotations.
• This method offers a promising solution for efficient and accurate cardiac structure segmentation in clinical practice.