A Dual-Task Synergy-Driven Generalization Framework for Pancreatic Cancer Segmentation in CT Scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Pancreatic cancer necessitates accurate lesion delineation for diagnosis and treatment.
• Existing segmentation methods struggle with imaging variability and lesion heterogeneity, limiting generalizability.
• Inter-patient variability and mimicry of normal tissues complicate accurate segmentation.

Purpose of the Study:

• To develop a generalized framework for accurate pancreatic lesion segmentation.
• To improve model stability and robustness across diverse imaging datasets.
• To enhance tumor localization and morphological characterization through integrated regression tasks.

Main Methods:

• A dual-task framework synergizing pixel-level classification and regression for lesion delineation.
• Integration of regression supervision within segmentation for enhanced generalization.
• Dual self-supervised learning in feature and output spaces for improved representational capability and stability.
• Reciprocal transformation of task outputs to bolster model performance.

Main Results:

• Achieved generalized pancreas segmentation comparable to in-domain performance (Dice: 84.07%) on three diverse datasets.
• Significantly improved cross-lesion generalized pancreatic cancer segmentation by 9.51%.
• Demonstrated enhanced model stability and representational capability across different imaging views.

Conclusions:

• The proposed dual-task framework offers resilient and efficient technological support for pancreatic cancer management.
• The model shows strong generalization capabilities, addressing limitations of current segmentation methods.
• This approach advances foundational technology for broader medical imaging applications.