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

Linkai Peng * 1 , Cuiling Sun * 1 , Zheyuan Zhang * 1 , Wanying Dou 1 , Halil Ertugrul Aktas 1 , Andrea M Bejar 1 , Elif Keles 1 , Tamas Gonda 2 , Michael B Wallace 3 , Zongwei Zhou 4 , Gorkem Durak 1 , Rajesh N Keswani 1 , Ulas Bagci † 1
1 Northwestern University , 2 New York University , 3 Mayo Clinic in Florida , 4 Johns Hopkins University

MIDL 2026

*Equal Contribution, †Corresponding Author

Paper Code arXiv Dataset

Abstract

Pancreas MRI segmentation remains challenging due to heterogeneous acquisition protocols and the limited availability of annotated multi-sequence data. While prior studies have examined cross-center robustness, the impact of variations across MRI sequences has received far less attention. We present CrossPan, a comprehensive empirical benchmark for evaluating segmentation robustness under heterogeneous abdominal MRI sequences. The benchmark comprises 1386 3D scans from eight institutions across three routinely acquired sequences (T1-weighted, T2-weighted, and Out-of-Phase). Across five tasks, we find that: (i) cross-sequence shifts constitute a substantially more severe challenge than cross-center variability; (ii) state-of-the-art DG methods show limited effectiveness in bridging the gap, while zero-shot foundation models exhibit heterogeneous robustness; and (iii) semi-supervised learning offers moderate gains at mid-range label ratios but remains unstable under extreme scarcity and exhibits strong sequence dependence. These results highlight the fundamental difficulty of cross-sequence generalization in abdominal MRI and position CrossPan as a rigorous platform for advancing robust and clinically reliable pancreas segmentation.