CrossPan: A Comprehensive Benchmark for Cross-Sequence Pancreas MRI Segmentation and Generalization

2026-04-22 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition, Medical Imaging · Depth: Expert, quick

Summary

CrossPan is a new multi-institutional benchmark designed to evaluate the generalization capabilities of deep learning models for pancreas MRI segmentation across different MRI sequences. The benchmark includes 1,386 3D scans from eight centers, covering T1-weighted, T2-weighted, and Out-of-Phase sequences. Experiments using CrossPan reveal that cross-sequence domain shifts pose a significantly greater challenge than cross-center variability; models achieving Dice scores above 0.85 in-domain drop to below 0.02 when transferred across sequences. Furthermore, traditional domain generalization methods offer minimal improvement, while foundation models like MedSAM2 show moderate zero-shot performance due to contrast-invariant shape priors. Semi-supervised learning is found to be effective only under stable intensity distributions, becoming unstable with high intra-organ variability. These findings highlight cross-sequence generalization as the primary obstacle to deploying pancreas MRI segmentation models clinically.

Key takeaway

For AI Scientists developing medical image segmentation models, you should prioritize addressing cross-sequence generalization over architectural improvements or increasing center diversity. Your models must maintain performance across diverse MRI sequences, as current methods fail catastrophically. Consider integrating foundation models that leverage contrast-invariant shape priors to improve robustness in zero-shot scenarios, and be wary of semi-supervised learning in highly variable sequences.

Key insights

Cross-sequence domain shift is the primary barrier to robust pancreas MRI segmentation.

Principles

• Cross-sequence shifts are more severe than cross-center variability.
• Foundation models use contrast-invariant shape priors for generalization.

In practice

• Prioritize cross-sequence generalization in model development.
• Evaluate foundation models for zero-shot segmentation tasks.

Topics

• Pancreas Segmentation
• Cross-Sequence Generalization
• MRI Domain Shift
• Deep Learning Models
• MedSAM2

Best for: AI Scientist, Research Scientist, Computer Vision Engineer

Related on AIssential

• CA-GCL: Cross-Anatomy Global-Local Contrastive Learning for Robust 3D Medical Image Understanding — CA-GCL uses global-local contrastive learning and clinical-aware text augmentation to prevent representation collapse in 3D medical FVLP.
• Mutual Distillation of Dual-Foundation Models for Semi-Supervised PET/CT Segmentation — Mutual distillation of dual-foundation models improves semi-supervised PET/CT segmentation with minimal labeled data.
• Pan-FM: A Pan-Organ Foundation Model with Saliency-Guided Masking for Missing Robustness — Pan-FM is a pan-organ foundation model robust to missing data via saliency-guided masking.
• Generative Deep Learning for Computational Destaining and Restaining of Unregistered Digital Pathology Images — Preprocessing-based domain adaptation enables cGAN models to generalize for computational pathology across institutions without retraining.
• A deep learning framework for glomeruli segmentation with boundary attention — A U-Net-based model with a boundary attention decoder improves glomeruli segmentation in kidney tissue.
• Multi-Contrast MRI Motion Correction via Parameter-Informed Disentanglement and Adaptive Experts — A unified MRI motion correction framework disentangles contrast and adapts to severity for robust, generalizable artifact removal.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.