Related Experiment Video
Updated: Jul 27, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
MSCDA: Multi-level semantic-guided contrast improves unsupervised domain adaptation for breast MRI segmentation in
Sheng Kuang1, Henry C Woodruff2, Renee Granzier3
1The D-Lab, Department of Precision Medicine, GROW - School or Oncology and Reproduction, Maastricht University, Maastricht, The Netherlands.
Summary
This study introduces a new unsupervised domain adaptation framework (MSCDA) for breast MRI segmentation, overcoming challenges from data variations. The method significantly improves segmentation accuracy by aligning image features across different domains.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Deep learning (DL) for breast MRI segmentation faces challenges due to domain shift from varying vendors, protocols, and patient heterogeneity.
- This domain shift hinders clinical implementation of automated segmentation tools.
Purpose of the Study:
- To propose a novel unsupervised Multi-level Semantic-guided Contrastive Domain Adaptation (MSCDA) framework.
- To address the domain shift problem in cross-domain breast MRI segmentation.
Main Methods:
- MSCDA integrates self-training with contrastive learning for unsupervised domain adaptation.
- It extends contrastive loss with multi-level contrasts (pixel-to-pixel, pixel-to-centroid, centroid-to-centroid) to leverage semantic information.
- A category-wise cross-domain sampling strategy and hybrid memory bank address data imbalance.
Main Results:
- MSCDA effectively aligns feature representations between source and target domains.
- The framework outperforms state-of-the-art methods in cross-domain breast MRI segmentation tasks.
- MSCDA demonstrates label efficiency, achieving strong performance with limited source data.
Conclusions:
- The proposed MSCDA framework offers a robust solution for unsupervised domain adaptation in medical image segmentation.
- It enhances feature alignment and segmentation accuracy, paving the way for improved clinical translation of DL models in breast MRI analysis.
