Generalizable Pancreas Segmentation via a Dual Self-Supervised Learning Framework

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Existing pancreas segmentation models perform well on single-source datasets but lack generalizability.
• Poor generalization leads to limited performance and instability when applied to data from different sources.
• Improving single-source generalization is crucial due to the scarcity of diverse medical imaging datasets.

Purpose of the Study:

• To enhance the generalization performance of pancreas segmentation models trained on single-source datasets.
• To develop a dual self-supervised learning approach leveraging both global and local anatomical contexts.
• To improve the characterization of high-uncertainty regions for more robust segmentation.

Main Methods:

• A dual self-supervised learning model incorporating global and local anatomical contexts.
• A global-feature contrastive self-supervised learning module guided by pancreatic spatial structure.
• A local-image-restoration self-supervised learning module to recover corrupted appearance patterns in high-uncertainty regions.

Main Results:

• State-of-the-art performance demonstrated on three pancreas datasets (467 cases).
• Comprehensive ablation analysis confirmed the effectiveness of the proposed method.
• Significant improvement in segmentation robustness and stability across different data sources.

Conclusions:

• The proposed dual self-supervised learning model effectively improves pancreas segmentation generalization.
• The method shows great potential for stable diagnostic and treatment support in pancreatic diseases.
• Exploiting global and local anatomical contexts enhances segmentation in challenging regions.