LMISA: A lightweight multi-modality image segmentation network via domain adaptation using gradient magnitude and...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Supervised medical image segmentation models struggle with cross-modality variations (e.g., CT vs. MRI).
• Intensity differences between imaging modalities significantly impact model performance, even for the same anatomical structures.
• Training data is often limited to a single modality, hindering generalization to others.

Purpose of the Study:

• To develop a novel end-to-end deep neural network for multi-modality medical image segmentation.
• To enable accurate segmentation in a target modality using labels only from a source modality (unsupervised domain adaptation).
• To address the challenge of intensity discrepancy and improve segmentation robustness across different imaging techniques.

Main Methods:

• A multi-resolution locally normalized gradient magnitude approach to minimize intensity differences between modalities.
• A dual-task encoder-decoder network for simultaneous image segmentation and reconstruction, facilitating domain adaptation.
• Adversarial learning to impose shape constraints and ensure consistent latent feature representation across domains.

Main Results:

• The proposed method achieves significantly higher performance in cross-modality segmentation (CT/MRI for kidney and cardiac tissues) compared to state-of-the-art approaches.
• Demonstrates superior segmentation results on unseen target domains without the need for model retraining.
• Exhibits lower model complexity while maintaining high accuracy.

Conclusions:

• The developed deep neural network effectively handles multi-modality medical image segmentation challenges.
• The approach offers robust and accurate segmentation for unlabeled target modalities, showing promise for clinical applications.
• The method provides a significant advancement in unsupervised domain adaptation for medical image analysis.