Contrastive Learning and Linear Self-Attention for Robust Multi-Center Pancreas Segmentation in MRI

Objectives: To develop a robust deep learning method for pancreas segmentation from multi-center MRI datasets by improving generalizability and accuracy. Method: A novel 3D segmentation model named CTraTL-UNet was implemented. The proposed method integrates linear self-attention and contrastive learning to enhance feature representation. It addresses inter-center heterogeneity of the pancreas dataset. The model was trained and validated using a multi-center and multi-modal (T1-weighted and T2-weighted MRI) dataset. The dataset is comprising 767 cases from five institutions. Findings: The proposed approach achieved superior segmentation performance with average Dice scores of 88.6% (T1W) and 89.2% (T2W).. The method demonstrated high reliability across varied institutional data, as confirmed by Jaccard, HD95, ASSD, and precision-recall metrics. Novelty: This work is the first to combine linear self-attention and instance-level contrastive loss in a 3D MRI segmentation method for pancreas. It is validated on a large-scale multi-center dataset and sets a new benchmark for cross-center robustness. Keywords: Pancreas segmentation, MRI, deep learning, linear self-attention, contrastive learning, multi-center dataset