Inter-subject registration-based one-shot segmentation with alternating union network for cardiac MRI images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning significantly aids disease diagnosis through medical image segmentation.
• Current methods often require large labeled datasets, which are time-consuming and costly to acquire.
• Limited labeled data poses a major challenge for training accurate medical image segmentation models.

Purpose of the Study:

• To propose a novel one-shot segmentation framework for cardiac MRI images.
• To address the challenge of limited labeled medical data by leveraging registration-based methods.
• To develop an efficient deep learning model that requires only a single labeled image for segmentation.

Main Methods:

• Introduced the Alternating Union Network (AUN), a one-shot segmentation framework utilizing inter-subject registration.
• Employed pre-processing steps including affine alignment and global intensity adjustment.
• Developed a two-subnetwork architecture trained alternately, incorporating global intensity similarity and intensity-independent structure registration.
• Defined a new Local Squared Error (LSE) similarity measurement for improved registration.

Main Results:

• The proposed framework successfully segments target cardiac MRI images using only one labeled source image.
• AUN leverages unlabeled data to enhance segmentation performance, demonstrating its clinical utility.
• The intensity-independent subnetwork and LSE measurement enable segmentation of images with complex intensity distributions.

Conclusions:

• The registration-based one-shot segmentation framework (AUN) effectively overcomes the scarcity of labeled medical images.
• AUN offers significant advantages for clinical applications by minimizing the need for extensive manual annotation.
• The developed method demonstrates robustness in segmenting medical images with challenging intensity variations.