PCLSegNet: Anatomy-Aware Deep Learning Framework for MRI-based Pancreatic Cyst Segmentation and Size Assessment
Authors: Hanyi Zhang, Tongyu Wang, Manuel Debic, Markus Wennmann, Johannes Haubold, Tamara Schikowski, Wilfried Reichardt, Fabian Bamberg, Christopher L. Schlett, Steffen Ringhof, Thoralf Niendorf, Tobias Pischon, Michael Leitzmann, Carsten Oliver Schmidt, Mark Oliver Wielpütz, Thomas Kroencke, Josua A. Decker, Stefan N. Willich, Thomas Keil, Rafael Mikolajczyk, Miriam Klauß, Hans-Ulrich Kauczor, Philipp Mayer, Bettina K. Budai
Organizations: Clinic for Diagnostic and Interventional Radiology, Heidelberg University Hospital, Medical Faculty of Heidelberg University, Heidelberg, Germany · Division of Radiology, German Cancer Research Center (DKFZ), Heidelberg, Germany · Department of Radiology, The Affiliated Hospital of Qingdao University, Qingdao, China · Division for Computational Radiology and Clinical AI, University Hospital Bonn, Faculty of Medicine at the University of Bonn, Bonn, Germany · Institute of Diagnostic and Interventional Radiology and Neuroradiology, University Hospital Essen, University Duisburg-Essen, Essen, Germany · IUF – Leibniz Research Institute for Environmental Medicine, Düsseldorf, Germany · Department of Diagnostic and Interventional Radiology, Medical Center – University of Freiburg, Faculty of Medicine – University of Freiburg, Freiburg, Germany · Max Delbrück Center for Molecular Medicine in the Helmholtz Association, Berlin-Buch, Germany · Institute for Epidemiology and Preventive Medicine, University of Regensburg, Regensburg, Germany · Institute of Diagnostic Radiology and Neuroradiology, University Medicine Greifswald, Greifswald, Germany · Clinic for Diagnostic and Interventional Radiology, University Hospital Augsburg, Augsburg, Germany · Institute of Social Medicine, Epidemiology and Health Economics, Charité – Universitätsmedizin Berlin, Berlin, Germany · Institute of Clinical Epidemiology and Biometry, University of Wuerzburg, Wuerzburg, Germany · State Institute of Health I, Bavarian Health and Food Safety Authority, Erlangen, Germany · Institute of Medical Epidemiology, Biometrics and Informatics, Medical Faculty of the Martin-Luther University Halle-Wittenberg, Halle (Saale), Germany
Early detection and follow-up of pancreatic cystic lesions (PCLs) are crucial, as some of their subtypes, such as mucinous cystic neoplasms (MCNs) and intraductal papillary mucinous neoplasms, carry a significant risk of malignant transformation. Although most AI studies investigating PCLs rely on computed tomography (CT) scans, magnetic resonance imaging (MRI) offers superior soft tissue contrast, making it highly suitable for PCLs assessment. In this work, we propose an anatomy-aware two-stage deep learning framework for automated PCLs segmentation on abdominal MRI scans. First, a publicly available pancreas segmentation model is used to localize the pancreas. Then, we introduce the anatomyaware PCLSegNet, which incorporates anatomical priors as guidance for pancreas and PCLs segmentation. To address the class imbalance problem, we develop a Class-Specific ErrorWeighted (CSEW) loss function. Postprocessing is applied through anatomical constraints.Our proposed method outperforms multi-organ segmentation models such as TotalSegmentator, achieving substantially higher Dice scores for pancreas segmentation. Moreover, it also achieves a superior normalized surface Dice (NSD) and recall for PCLs segmentationcompared to other state-of-the-art models like nnU-Net, while also notably reducing the false negative rate, facilitating the automatic early detection and the noninvasive imaging-based assessment of PCLs.
Figures & tables
Figure 1: In stage 1, a 3D abdominal MRI scan is processed using PanSegNet to segment the pancreas, from which a bounding box is extracted and enlarged to form a region of interest (ROI). This ROI, along with the pancreas mask, is used to generate stacked 2.5D inputs.
Figure 2: In stage 2, the pancreas ROI is processed by an image encoder, while the anatomical prior (predicted pancreas mask in stage 1) is passed through a convolutional block followed by interpolation. Both features are fused in the bottleneck using an Attention Fusion Gate and then decoded to generate the final segmentations. The model is supervised using the Class-Specific Error-Weighted loss. In the postprocessing step, predicted PCLs are refined using spatial anatomical constraints derived from multi-organ segmentation, which removes false positives located far from the pancreas or in adjacent organs.
Characteristics
Training dataset ( n=900 )
Test dataset ( n=100 )
Sex, n (%)
Female
453 (50.3)
50 (50.0)
Male
447 (49.7)
50 (50.0)
Age, years
56 [48–63]
56.5 [49–64]
Largest cyst size per participant, mm
7 [5–12]
7 [5–10]
Largest cyst location, n (%)
Table 1: Demographic and PCL characteristics in training and test datasets. Median [interquartile range, 25th–75th percentile] values are reported for continuous variables.
Model usage
Method
Param (M)
Input scope
DSC
Zero-shot
TotalSegmentator MRI
176.50
Full 3D scan
0.49
MRSegmentator
88.20
Full 3D scan
0.72
PanSegNet
43.90
Full 3D scan
0.91
In-house trained
2D nnU-Net
31.00
2D ROI
0.89
3D nnU-Net
41.20
3D ROI
0.90
Attention U-Net
23.40
3D ROI
0.83 ± 0.07
Table 2: Performance comparison of different models for pancreas segmentation (mean DSC ± standard deviation). Models are grouped by usage: zero-shot versus in-house trained models. Model size is reported as the total number of parameters in millions (M). Best results are shown in bold and second-best results are underlined.
Figure 3: Qualitative comparison of pancreas segmentation results on an example test case. For visual demonstration of the pancreas segmentation performance, the segmentation results of the nine trained deep learning models are shown together with the ground truth mask annotated by the radiologist on a representative example case. While the basic shape and location of the pancreas is captured on all predicted pancreas masks in this example, the predictions of TotalSegmentator MRI and MRSegmentator constantly appeared smooth, allowing for only a coarse segmentation and losing the texture details. In comparison, both 2D and 3D nnU-Net, as well as VNet and our proposed PCLSegNet model, achieved superior results by better capturing surface granularity. Although Attention U-Net and UNETR achieved relatively good Dice scores, they noticeably predicted false positive regions, leading to less accurate results. Rows from top to bottom: axial slices, coronal slices, and 3D surface reconstructions with Dice similarity coefficients.
PCL-level metrics
Method
Dice ↑
NSD ↑
Inv. HD95 ↑
Recall ↑
TP (#) ↑
FN (#) ↓
Size group ≤ 5 mm (total PCLs: 13)
2D nnU-Net
0.25
0.29
0.12
0.31
4
9
3D nnU-Net
0.14
0.20
0.09
0.23
3
10
Attention U-Net
0.09 ± 0.02
0.13 ± 0.01
0.06 ± 0.01
0.15 ± 0.00
2.00 ± 0.00
11.00 ± 0.00
UNETR
0.26 ± 0.05
0.41 ± 0.09
0.19 ± 0.03
0.51 ± 0.08
6.60 ± 1.02
6.40 ± 1.02
Table 4: PCL-level segmentation performance across different PCL size groups (Kyoto classification). Results are presented as mean ± standard deviation where applicable. Best results are shown in bold and second-best results are underlined.
Figure 4: Visual comparison of segmentation results. Rows show the largest cross-section of the original MRI image, the ground truth (GT) segmentation, and the predictions of the five state-of-the-art models versus our proposed PCLSegNet. Columns (a)–(d) represent four PCLs from diameter groups with different clinical relevance (classified according to the Kyoto guidelines into ≤ 5 mm vs. 5–10 mm vs. 10–20 mm vs. > 20 mm). The normalized surface Dice (NSD) score is displayed in the bottom-right corner. Please note that the images show the largest cross-sectional axial slice, but the metrics are calculated from 3D segmentations. Column (a) represents a PCL from the size-group ≤ 5 mm located in the pancreatic body. Notably, both the 3D nnU-Net and Attention U-Net failed to segment this PCL, misclassifying it as pancreas parenchyma. While the 2D nnU-Net and V-Net achieved an NSD of 0.79, PCLSegNet achieved 0.98, demonstrating near-perfect segmentation. A similar trend is evident in column (b) (size-group > 5 mm to ≤ 10 mm), where PCLSegNet maintains strong performance with an NSD of 0.90, whereas the 2D nnU-Net, Attention U-Net, and V-Net all failed to detect or segment the target PCL. For larger PCLs, as shown in columns (c) and (d), all models successfully detected the lesions, but with varying degrees of segmentation accuracy. In column (c), the PCL was measured 16 mm large by the radiologist. The largest diameter measured by PCLSegNet was 17 mm, while Attention U-Net and 2D nnU-Net measured 16 mm and 17 mm, respectively, showing close agreement with the ground truth. Conversely, 3D nnU-Net and V-Net yielded less accurate measurements of 19 mm and 12 mm, respectively, despite their relatively high NSD scores. In example (d), featuring an exceptionally large PCL (maximum diameter of 44 mm), PCLSegNet achieved a perfect segmentation with an NSD of 1.00, demonstrating superior volumetric capture compared to all baseline models.
Model
Loss
Postproc.
DSC
NSD
Inv. HD95
Precision
Recall
F1
U-Net
DiceCE
No
0.39 ± 0.01
0.52 ± 0.02
0.18 ± 0.01
0.80 ± 0.04
0.63 ± 0.02
0.70 ± 0.02
PCLSegNet
CSEW
No
0.52 ± 0.02
0.69 ± 0.03
0.24 ± 0.00
0.67 ± 0.07
0.78 ± 0.03
0.72 ± 0.03
PCLSegNet
CSEW
Yes
0.52 ± 0.02
0.68 ± 0.03
0.23 ± 0.00
0.74 ± 0.06
0.77 ± 0.03
0.75 ± 0.02
Table 5: Ablation study. Starting from a baseline U-Net trained with DiceCE, the effects of the anatomy-aware architecture, CSEW, and postprocessing strategy were evaluated. Results are presented as mean ± standard deviation. Best results are shown in bold and second-best results are underlined.
Maximum diameter
Size-based grouping ∗
Model
CCC [95% CI]
Weighted κ [95% CI]
2D nnU-Net
0.92 [0.58, 0.97]
0.64 [0.43, 0.81]
3D nnU-Net
0.93 [0.56, 0.97]
0.67 [0.43, 0.82]
Attention U-Net
0.72 [0.42, 0.79]
0.48 [0.18, 0.71]
UNETR
0.15 [ − 0.01, 0.43]
0.28 [0.07, 0.49]
V-Net
0.64 [0.38, 0.73]
0.48 [0.15, 0.72]
Table 6: Agreement analysis of diameter measurement and size-based classification: Lin’s concordance correlation coefficient (CCC) and quadratic weighted κ with their respective 95% confidence intervals [95% CI]. Best results are shown in bold and second-best results are underlined.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Algorithm S1: Greedy matching for PCL instance detection.
1:
Initialize Mgt←∅ and Mpred←∅
2:
for each ground truth PCL cgt in GT do
3:
Find the predicted PCL cpred with the largest overlap with cgt
Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.
Background: Limited tumor annotations and incomplete labels across MRI datasets hinder unified segmentation of pancreatic tumors and organs at risk (OARs) for MRI-guided radiotherapy. Purpose: We propose Pancreas-MedNeXt, a unified model that segments the whole pancreas, its subregions, tumors, and surrounding OARs from one MRI volume. Methods: The model uses a shared MedNeXt encoder-decoder with supervised whole-pancreas and subregion pretraining, followed by multitask fine-tuning. Three output heads predict the pancreatic subregions, tumor, and OARs (stomach, duodenum, kidneys). An annotation-aware loss allows training on partially labeled data without treating unannotated structures as background. Pancreas and tumor were evaluated by three-fold cross-validation on 142 PANTHER patients (92 diagnostic, 50 MR-Linac), subregions and OARs on Cyst-X and AMOS test sets, and pancreas and OARs on 33 external T1-weighted TotalSegmentator scans. Comparisons with nnU-Net, MedNeXt, VISTA3D, and Swin UNETR used the Dice similarity coefficient (DSC) with paired bootstrap 95% CIs. Results: Pancreas-MedNeXt achieved the highest mean pancreas and tumor DSC on diagnostic and MR-Linac MRI. Tumor DSC was 58.41% and 42.55%, exceeding task-specific MedNeXt by 4.94 (95% CI: 0.28-9.45) and 5.89 (95% CI: 2.37-9.72) percentage points, respectively. Whole-pancreas DSC was 85.73% and 70.66%. Pretraining improved head, body, and tail DSC by 2.01, 3.00, and 3.78 percentage points over training from scratch. It also achieved the highest DSC for the pancreas, duodenum, and right kidney on AMOS, and for all five organs on external MRI, exceeding the next-best method by 4.92-8.85 percentage points.Conclusion: A single Pancreas-MedNeXt model segments tumors, pancreatic subregions, and OARs on diagnostic and MR-Linac MRI, which could streamline contouring for pancreatic cancer radiotherapy.
Zach Eidex, Mojtaba Safari, Kundan Thind +3
Department of Computer Science and Radiation Oncology, Emory University, Atlanta, GA, USA · Department of Radiation and Cellular Oncology, University of Chicago, Chicago, IL 60637, USA · Department of Radiation Oncology, Henry Ford Health, Detroit, MI, USA
Automatic pancreas segmentation is fundamental to abdominal MRI analysis, yet deep learning models trained on one MRI sequence often fail catastrophically when applied to another-a challenge that has received little systematic investigation. We introduce CrossPan, a multi-institutional benchmark comprising 1,386 3D scans across three routinely acquired sequences (T1-weighted, T2-weighted, and Out-of-Phase) from eight centers. Our experiments reveal three key findings. First, cross-sequence domain shifts are far more severe than cross-center variability: models achieving Dice scores above 0.85 in-domain collapse to near-zero (<0.02) when transferred across sequences. Second, state-of-the-art domain generalization methods provide negligible benefit under these physics-driven contrast inversions, whereas foundation models like MedSAM2 maintain moderate zero-shot performance through contrast-invariant shape priors. Third, semi-supervised learning offers gains only under stable intensity distributions and becomes unstable on sequences with high intra-organ variability. These results establish cross-sequence generalization-not model architecture or center diversity-as the primary barrier to clinically deployable pancreas MRI segmentation. Dataset and code are available at https://crosspan.netlify.app/.
Linkai Peng, Cuiling Sun, Zheyuan Zhang +10
Department of Radiology, Northwestern University, Chicago, USA. · Department of Electrical and Computer Engineering, Northwestern University, Evanston, USA. · Department of Computer Science, Northwestern University, Evanston, USA. +4
Intraductal papillary mucinous neoplasms (IPMNs) are pancreatic cancer precursors, and guideline-based risk stratification both over-treats indolent lesions and misses progression. We present Cyst-X, a public benchmark of 1,461 T1-weighted (T1W) and T2-weighted (T2W) MRI volumes from 764 patients at seven centers, with expert pancreas masks, risk labels, splits and reference AI risk-prediction models. PanSegNet segmented the pancreas with Dice of 86.34% (T1W) and 89.26% (T2W) on held-out centers. For the primary task, high-risk versus low-risk IPMN, a 3D DenseNet-121 achieved cross-validated AUCs of 74.59% and 73.45%; complementary broad binary and three-class tasks reached 78.67--81.28% and 73.47--75.76%. Because patient images often cannot be pooled across institutions, we also simulated federated learning across the seven centers. On T2W MRI, FedProx was similar to centralized training in center-averaged external AUC (71.00% versus 71.18%) but not in pooled AUC (62.35% versus 77.49%). Radiologist comparisons showed cohort-dependent sensitivity--specificity tradeoffs. Cyst-X supports reproducible method development and identifies within-IPMN risk discrimination and cross-center score comparability as open challenges.
Hongyi Pan, Gorkem Durak, Halil Ertugrul Aktas +27
Machine & Hybrid Intelligence Lab, Department of Radiology, Northwestern University, Chicago, IL, USA. · Istanbul Faculty of Medicine, Istanbul University, Istanbul, Turkey. · Department of Preventive Medicine, Northwestern University, Chicago, IL, USA. +10