Unsupervised Bidirectional Cross-Modality Adaptation via Deeply Synergistic Image and Feature Alignment for Medical...

Area of Science:

• Medical Image Computing
• Artificial Intelligence in Medicine
• Deep Learning

Background:

• Deep neural networks (DNNs) in medical imaging suffer performance degradation when applied to new datasets with different characteristics.
• Unsupervised domain adaptation (UDA) is crucial for adapting DNNs to unlabeled target domains, addressing data heterogeneity.

Purpose of the Study:

• To develop a novel UDA framework, Synergistic Image and Feature Alignment (SIFA), for medical image segmentation.
• To effectively adapt segmentation networks to unlabeled target domains by aligning image appearance and feature representations.

Main Methods:

• SIFA employs adversarial learning for simultaneous image appearance transformation and domain-invariant feature enhancement.
• A deeply supervised mechanism and a shared feature encoder are utilized for synergistic adaptation.
• The framework is trained end-to-end to maximize mutual benefits between image and feature alignment.

Main Results:

• SIFA demonstrated effectiveness in cardiac substructure and abdominal multi-organ segmentation tasks.
• Bidirectional cross-modality adaptation between MRI and CT images was successfully achieved.
• The proposed SIFA method significantly outperformed existing state-of-the-art domain adaptation approaches.

Conclusions:

• SIFA provides a powerful and effective solution for unsupervised domain adaptation in medical image segmentation.
• The synergistic alignment of image and feature domains enhances model robustness and generalization to unseen data.
• This framework holds significant potential for improving the clinical applicability of deep learning models in medical imaging.