Source-linked AI summary
Uncertainty-aware Coarse-to-Fine Main Pancreatic Duct Segmentation for Width Measurement in 3D MRCP
Zipei Zhang, Jaakko Sahlsten, Mikko Saukkoriipi, Johanna Ronkainen, Mona Lehto, Akseli Aittoniemi, Irina Rinta-Kiikka, Kimmo Kaski
TL;DR
Fully automated MPD segmentation and width analysis from MRCP remains an unmet research need, despite the relevance of duct morphology to pancreatic cancer assessment. This study uses a coarse-to-fine pipeline with ensemble-based uncertainty analysis; it reports sub-millimeter width error and a relationship between uncertainty and segmentation performance.
Problem
Fully automated MRCP-based MPD segmentation and width analysis has not been established, despite duct width’s relevance to pancreatic cancer assessment.
Source: §1 · p. 2Method
A coarse model localises a high-recall duct region, then a fine model segments it, with ensemble-based uncertainty analysis.
Source: §2.1 · p. 3, §2.2 · p. 3, §4 · p. 10Results
Maximum-width error was sub-millimeter, and higher uncertainty correlated with lower segmentation performance.
Source: §4 · p. 10Takeaways & Limitations
Model-derived width may support screening or prioritisation, but false-positive dilatation predictions still require radiologist review.
Source: §4 · p. 10The test set comprised 14 patients and 55 scans from one hospital, limiting generalizability.
Source: §4 · p. 10
Abstract
from arXiv · showhide
Assessment of the main pancreatic duct (MPD) from volumetric magnetic resonance cholangiopancreatography (MRCP) scans is an essential task for ductal dilatation diagnosis and pancreatic cancer analysis. However, manual assessment is affected by reader variability in which automated analysis may assist. Nevertheless, fully automated and uncertainty-aware MPD analysis from volumetric MRCP remains underexplored area of research. In this study, we propose an automatic and uncertainty-aware deep learning-based pipeline for MPD segmentation and maximum-width analysis from MRCP scans. A coarse localisation model first identifies the ductal region-of-interest, followed by a fine model that segments the MPD within the region. The dataset consists of 256 MPD-segmented MRCP scans from 69 patients. A randomly selected holdout test subset of 14 patients with 55 MRCP volumes was used for evaluation. We employed a five-model cross-validation ensemble for improved performance and uncertainty estimation. The best MPD segmentation performance was 0.634 DSC, 0.744 precision, and 0.593 recall. The mean absolute error for maximum MPD width prediction was 0.355~mm and dilated MPD ($\geq 5$~mm) was detected with 75.0\% sensitivity and 74.4\% specificity. These novel results suggest that automated MPD segmentation can be used to aid MRCP-based duct measurements.
1 Introduction
The introduction motivates automated MPD assessment by its relevance to pancreatic cancer and the importance of duct morphology. It identifies a gap in fully automated MPD segmentation and width analysis from MRCP.
1 Introduction
- More than 80% of patients with pancreatic ductal adenocarcinoma are diagnosed late, making early detection important; imaging can reveal tumors and secondary duct changes.The introduction notes that tumors may become inoperable after infiltrating adjacent structures.Source: §1 · p. 1
- MRCP highlights fluid-filled ducts, and MPD dilatation can be a secondary imaging finding associated with an obstructing tumour.The cited background describes abrupt strictures over 6 mm in malignant cases.Source: §1 · p. 1
- Focal MPD diameter change had the shortest duration from development to diagnosis in a cited study [7], while duct volume and diameter improved PDAC classification over radiomics alone [9].Source: §1 · p. 1
- Prior automatic MPD analyses were largely CT-based [14–16], while MRCP work included manual or semi-automatic segmentation and measurement-focused studies [17–20].The introduction identifies the absence of a fully automated MRCP MPD segmentation-and-width analysis pipeline as the motivating gap.Source: §1 · p. 2
2 Methods
The study evaluates a coarse-to-fine pipeline for MPD segmentation and maximum-width analysis in 3D MRCP, using a patient-level dataset split and expert-reviewed annotations. Five-fold model ensembles provide final predictions and uncertainty estimates, which are assessed alongside segmentation and width-measurement performance.
2.1 Dataset, annotation and preprocessing
- The dataset contains 256 retrospective MRCP volumes from 69 patients, with a patient-level 80:20 development/test split yielding 201 and 55 scans.The test set comprises 55 scans from 14 patients; two clinical researchers created voxel-wise annotations that two experienced abdominal radiologists reviewed.Source: §2.1 · pp. 2–3
2.2 Segmentation models
- A 3D U-Net served as the coarse model and a 3D SegResNet as the fine model; established architectures were compared for practical feasibility and validation performance.Selection criteria also included GPU memory demand and convergence stability.Source: §2.2 · p. 3
- The pipeline localizes MPD on a downsampled full volume, expands the resulting ROI with a safety margin, then segments the full-resolution crop with a fine model.The final ROI prediction is mapped back to the original image space.Source: §2.2 · pp. 3–4
2.3 Implementation details
- Both segmentation stages used weighted cross-entropy and Dice loss, with a larger class weight for rare MPD voxels; training also used 3D augmentations and connected-component post-processing.Post-processing removes small isolated regions not connected to the largest component.Source: §2.3 · pp. 4–5
- Five patient-level cross-validation model pairs were applied independently at test time, and their probability maps were averaged for the final ensemble segmentation.The same ensemble predictions were used for uncertainty analysis.Source: §2.3 · p. 4
2.4 Evaluation metrics
- Uncertainty was evaluated using volume-level predictive-entropy aggregations and coefficient of variation, with referral-based and correlation analyses against model performance.Evaluation also covered volumetric and surface-distance segmentation metrics, radiologist review, maximum-width error, and classification at a 5 mm dilation threshold.Source: §2.3 · p. 5, §2.4 · pp. 5–6
3 Results
The 3D U-Net + 3D SegResNet pipeline had the strongest reported test-set segmentation metrics, and its width measurements showed moderate performance. The uncertainty analyses related model uncertainty to segmentation performance and examined variation across patients.
3.1 Segmentation performance
- 3D U-Net + 3D SegResNet achieved the highest Dice similarity coefficient, precision, and recall among the four test-set pipelines.Table 2 summarizes performance on 55 test-set volumes.Source: §3.1 · p. 6, Table 2 · p. 6
3.2 Qualitative evaluation
- The model’s maximum-width measurements had 0.355 mm mean absolute error, with 75.0% sensitivity and 74.4% specificity for detecting dilated MPD.Dilatation was defined as maximum width ≥5 mm.Source: §3.2 · p. 6
3.3 Model Uncertainty Analysis
- At 30% data remaining, the best model’s DSC was 0.730 with CV referral, 0.677 with SH, and 0.665 with H.For CV referral, the best model’s DSC rose from 0.634 at 100% data remaining; Figure 1 presents the referral-threshold evaluations.Source: §3.3 · p. 7, Figure 1 · p. 7
- CV correlated with DSC (r = −0.759), ASSD (r = 0.693), and HD95 (r = 0.695), each at p < 0.001.Its correlation with MPD maximum-width MAE was not significant (ρ = 0.187, p = 0.173); Figure 2 shows quantitative and qualitative comparisons.Source: §3.3 · p. 7, Figure 2 · p. 8
3.4 Within-patient analysis
- Within-patient mean distances from the patient mean were 0.05 (0.04) for DSC and 0.16 (0.32) for MPD maximum-width MAE.Figure 3 shows patient-level segmentation, radiologist qualitative measures, and CV, with patients ordered by median DSC.Source: §3.4 · p. 8, Figure 3 · p. 9
4 Discussion
The discussion reports that uncertainty estimates track segmentation quality, while model-derived width measurements may support screening but still require radiologist review. The authors also note variable segmentation failures and a small, single-hospital test set, underscoring the need for external validation and robustness improvements.
4 Discussion
- The 3D U-Net + 3D SegResNet approach outperformed the other segmentation methods, despite producing one empty prediction mask and the second-lowest ASSD.The authors attribute its performance to residual feature extraction that may stabilize optimization and preserve local duct boundaries on a limited dataset.Source: §4 · p. 10
- Model-derived maximum width may support screening or prioritisation, but false-positive dilatation predictions require radiologist review.The automatic approach performed below inter-rater agreement, consistent with segmentation metrics that still leave room for improvement.Source: §4 · p. 10
- Visual usability ratings were 2.67 and 2.50, and errors included a complete duct miss; the authors suggest data-distribution shifts may affect generalisability.Despite imperfect anatomical masks, sub-millimeter maximum-width error may still support measurement when maximum width, rather than complete duct delineation, is the endpoint.Source: §4 · p. 10
- 74.5% accuracy for DMPD classification was below the raters’ 83.6% agreement; coefficient of variation outperformed entropy-based uncertainty measures.Uncertainty estimates correlated with segmentation quality, and the 3D U-Net + SegResNet3D model performed best in referral-based analysis.Source: §4 · p. 10
- The test set comprised 14 patients and 55 scans from one hospital, while reference measurements were limited by image resolution and observer judgment.The authors advise assuming some measurement uncertainty when interpreting reported distances.Source: §4 · p. 10
- The authors call for external validation, a larger dataset, better MPD continuity and morphology preservation, and more robust automatic width estimation.They suggest orthogonal cross-sectional analysis as a way to reduce dependence on manual expert measurement.Source: §4 · p. 10