Moment-Consistent Contrastive CycleGAN for Cross-Domain Pancreatic Image Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate pancreas segmentation in CT and MR images is crucial for pancreatic cancer diagnosis and treatment.
• Traditional supervised methods demand extensive labeled data, which is time-consuming and laborious.
• Domain shift hinders the deployment of segmentation networks across different imaging modalities.

Purpose of the Study:

• To develop a cross-domain pancreas segmentation algorithm to overcome limitations of traditional methods.
• To improve the accuracy and robustness of pancreas segmentation across different imaging modalities.
• To reduce the need for large amounts of labeled training data in medical image segmentation.

Main Methods:

• Proposed a Moment-Consistent Contrastive Cycle Generative Adversarial Networks (MC-CCycleGAN) for style transfer and feature extraction.
• Introduced multi-order central moments and contrastive loss to maintain pancreatic structure and shape consistency during style transfer.
• Implemented a multi-teacher knowledge distillation framework to enhance student network performance and robustness.

Main Results:

• The MC-CCycleGAN effectively extracts structural features while eliminating redundant style features during style transfer.
• Moment consistency constraints successfully preserved pancreatic anatomy before and after style transfer.
• The proposed framework demonstrated superior performance compared to state-of-the-art domain adaptation methods in experimental evaluations.

Conclusions:

• The developed cross-domain segmentation algorithm significantly advances pancreas segmentation in medical imaging.
• MC-CCycleGAN offers an effective solution for domain adaptation in medical image segmentation tasks.
• The study highlights the potential of knowledge distillation and moment consistency for robust and accurate segmentation.