Contrast-Invariant Self-supervised Segmentation for Quantitative Placental MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Biomedical Image Analysis

Background:

• Accurate placental segmentation is vital for quantitative analysis in T2*-weighted MRI.
• Challenges include echo-dependent contrast variation and limited manual annotations.
• Existing methods struggle with robustness across different echo times.

Purpose of the Study:

• To develop a robust segmentation framework for multi-echo T2*-weighted placental MRI.
• To address challenges posed by contrast variation and limited annotations.
• To improve quantitative analysis through contrast-invariant representations.

Main Methods:

• A contrast-augmented segmentation framework integrating masked autoencoding (MAE) for self-supervised pretraining.
• Masked pseudo-labeling (MPL) for semi-supervised domain adaptation across echo times.
• Global-local collaboration and a semantic matching loss for representation consistency.

Main Results:

• The proposed framework demonstrates effective generalization across echo times.
• Outperforms traditional supervised segmentation baselines on a clinical dataset.
• Achieves robust, contrast-invariant representations for placental segmentation.

Conclusions:

• This work presents the first systematic framework for multi-echo T2*-weighted placental MRI segmentation.
• The approach enhances segmentation accuracy and robustness in challenging MRI conditions.
• Enables more reliable quantitative analysis of placental structures.