PCLSegNet: Anatomy-Aware Deep Learning Framework for MRI-based Pancreatic Cyst Segmentation and Size Assessment

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. BudaiarXiv 2026 · arXiv

TL;DR

Researchers have developed PCLSegNet, a two-stage deep learning framework that uses anatomical priors to automate the segmentation and size assessment of pancreatic cystic lesions (PCLs) on MRI. By integrating a coarse pancreas mask as guidance and employing a specialized loss function, the model achieves state-of-the-art detection sensitivity while remaining significantly more lightweight than existing heavy-duty architectures.

What Problem Does This Paper Address?

Pancreatic cystic lesions (PCLs) are common incidental findings on abdominal imaging, but they aren't just "noise"—certain subtypes carry a high risk of turning into pancreatic cancer. Early detection and precise monitoring of their size are critical for patient management.

While Magnetic Resonance Imaging (MRI) is the gold standard for assessing these cysts due to its superior soft-tissue contrast, radiologists often face significant challenges:

  1. High Interobserver Variability: Different experts often disagree on exact size measurements.
  2. Under-reporting: Up to two-thirds of incidental PCLs are missed or omitted in routine clinical reports.
  3. Anatomical Complexity: The pancreas varies wildly in shape, and cysts can be tiny, making automated segmentation a "needle in a haystack" problem.

Existing AI tools mostly focus on CT scans or require massive computational resources. This paper introduces a specialized, efficient pipeline tailored specifically for the nuances of MRI-based PCL assessment.

Methods

The authors propose a two-stage "anatomy-aware" framework:

  1. Stage 1: Pancreas Localization: A pre-trained model (PanSegNet) identifies the pancreas in a full 3D MRI scan to define a Region of Interest (ROI). This drastically reduces the search space for the next step.
  2. Stage 2: Anatomy-Aware PCL Segmentation: This is the core contribution. The researchers used a custom 2.5D U-Net (processing stacks of three consecutive axial slices).
  3. Anatomical Prior: The coarse pancreas mask from Stage 1 is transformed into a spatial feature map and fused with the image data using an Attention Fusion Gate. This "teaches" the model to prioritize the pancreatic parenchyma and ignore noise from adjacent organs.
  4. CSEW Loss: To handle the extreme class imbalance (tiny cysts vs. large background), they developed a Class-Specific Error-Weighted (CSEW) loss function that dynamically penalizes errors on the PCL class.
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.

Finally, a postprocessing step uses anatomical constraints from multi-organ segmentation to filter out false positives located far from the pancreas.

Datasets

The study utilized a large-scale dataset from the German National Cohort (NAKO):

  • Total Participants: 1,000 randomly selected individuals.
  • Training Set: 900 participants (450 with cysts, 450 without).
  • Test Set: 100 participants (50 with cysts, 50 without).
  • Ground Truth: All scans were manually annotated by three experienced radiologists to ensure a high-quality reference standard.

Key Insights and Ideas

  • Anatomy as a "North Star": The most significant "aha" moment is that providing the model with a coarse map of where the pancreas is (the anatomical prior) significantly boosts performance. It prevents the model from hallucinating cysts in the duodenum or bile ducts.
  • Efficiency Matters: By using a 2.5D approach and a specialized architecture, PCLSegNet uses only 8.2 million parameters. For context, that's about 1/15th the size of the popular UNETR transformer model, yet it delivers better clinical results.
  • Beyond Dice Scores: The authors emphasize the Normalized Surface Dice (NSD) metric. Traditional voxel-based metrics like the Dice Similarity Coefficient (DSC) can be overly sensitive to minor boundary deviations in tiny objects. NSD provides a more clinically relevant measure of how well the model captures the cyst's boundary.

Key Results

PCLSegNet outperformed several state-of-the-art (SOTA) models, including 2D and 3D versions of the industry-standard nnU-Net:

  • Detection Sensitivity: Achieved a True Positive Rate (TPR) of 84.0% at the participant level, a 6% improvement over the next best model.
  • Segmentation Accuracy: At the individual cyst level, it reached a DSC of 0.52 and an NSD of 0.68, consistently outperforming heavier models across all cyst size categories.
  • Size Assessment: The automated diameter measurements showed "nearly perfect" agreement with radiologists, with a Concordance Correlation Coefficient (CCC) of 0.97.
  • Small Lesion Performance: Even for the most challenging cysts (≤ 5 mm), PCLSegNet maintained superior performance where other models often failed entirely.
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.

Limitations and Future Work

While promising, the authors note a few areas for growth:

  • Sequence Specificity: The model was trained and validated only on T2-weighted HASTE sequences. Its performance on other common MRI sequences (like T1-weighted or fat-suppressed scans) remains to be tested.
  • No Subtyping: The current framework detects and measures cysts but doesn't yet classify them into specific pathological subtypes (e.g., distinguishing a benign cyst from a high-risk one).
  • Future Directions: The team plans to incorporate more anatomical context (like the bile and pancreatic ducts) and extend the pipeline to support longitudinal tracking—automatically comparing size changes across multiple scans over time.

Summary updated

Abstract

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.
BibTeX
@misc{zhang2026pclsegnet,
  author = {Hanyi Zhang and Tongyu Wang and Manuel Debic and Markus Wennmann and Johannes Haubold and Tamara Schikowski and Wilfried Reichardt and Fabian Bamberg and Christopher L. Schlett and Steffen Ringhof and Thoralf Niendorf and Tobias Pischon and Michael Leitzmann and Carsten Oliver Schmidt and Mark Oliver Wielpütz and Thomas Kroencke and Josua A. Decker and Stefan N. Willich and Thomas Keil and Rafael Mikolajczyk and Miriam Klauß and Hans-Ulrich Kauczor and Philipp Mayer and Bettina K. Budai},
  title = {{PCLSegNet: Anatomy-Aware Deep Learning Framework for MRI-based Pancreatic Cyst Segmentation and Size Assessment}},
  year = {2026},
  eprint = {2611.15278},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15278},
}
Figure