REVIEW 5 major objections 6 minor 6 references
Pediatric Pancreas Segmentation from MRI Scans with Deep Learning
T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read A deep-learning pancreas segmentation model trained only on adult MRI transfers to children aged 2–19, reaching Dice scores of 0.88 in healthy controls and 0.80–0.81 in pancreatitis without pediatric retraining.
desk verdict A useful public pediatric MRI dataset and honest external validation, but the 'expert-level' claim is not supported by the paper's own inter-observer numbers. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is PanSegNet: a volumetric nnU-Net whose convolutional encoder-decoder is augmented with a linear self-attention transformer (a self-attention mechanism that avoids the quadratic cost of standard attention), trained on 767 adult MRI scans from 499 participants. The self-attention component lets the model weigh long-range context across the whole 3D volume while keeping computation feasible, and the adult training supplies the learned pancreatic appearance priors. Because the model is applied to the pediatric cohort without fine-tuning, the pediatric scans function as a pure external generalization test, and the paper's argument rests on that transfer.
What would settle it
Run PanSegNet on an independent, multi-institution pediatric MRI cohort with separate expert consensus labels and compare it against the same four benchmarks; if its pancreatitis Dice falls below 0.75 or is statistically indistinguishable from nnUNet3D/TransUNet, the claim of state-of-the-art transferability would be falsified.
Extended reading notes
Core claim
The central claim is that adult-trained pancreas segmentation transfers to pediatric anatomy and to inflammatory disease states without adaptation. PanSegNet, a 3D nnU-Net augmented with a linear self-attention module, was trained on 767 adult MRI scans and then frozen; the 84 pediatric scans were used only for external testing. On this test set it reports an overall Dice of 0.85, with group-wise scores of 0.88 (healthy), 0.81 (acute pancreatitis), and 0.80 (chronic pancreatitis), and it outperforms four benchmark architectures on Dice and Jaccard index, while nnUNet3D records a marginally smaller HD95. The paper also reports that human inter-observer Dice for the same reference segmentations is 0.82–0.86, and that automated volumes correlate with manual volumes at $R^2=0.85$ in healthy children and $R^2=0.77$ in diseased children, which it reads as evidence of clinical reliability for volume-based follow-up.
Load-bearing premise
The load-bearing premise is that the radiologists' manual segmentations are the true pancreas boundary even though two readers agree at only Dice 0.82–0.86, so the model's 0.80–0.88 scores could partly reflect label noise rather than true anatomical accuracy.
Editorial extensions
If this is right
- Automatic pancreas volume from routine T2 MRI becomes feasible in children, with $R^2=0.85$ in healthy and $R^2=0.77$ in diseased cases relative to manual volumetry, supporting non-invasive longitudinal monitoring.
- Because the model was not retrained, adult-trained pancreas segmentation can be transferred to pediatric populations, reducing the need for large pediatric labeled datasets.
- The released annotated dataset and code give other groups a benchmark for pediatric pancreas MRI segmentation, enabling direct comparisons and further training.
- In acute and chronic pancreatitis, lower Dice and higher HD95 (9.85 mm and 15.67 mm) mark fluid collections and fibrotic atrophy as the main failure modes, guiding future model improvements such as fluid-aware augmentation.
- Benchmark results suggest that single-organ 3D approaches with self-attention outperform generic multi-organ or 2D pipelines on this task, although nnUNet3D is competitive on boundary distance.
Reading between the lines
- Editorial inference: because the model's Dice gaps across groups (0.88 versus 0.81/0.80) are comparable to inter-observer Dice differences (0.86 versus 0.82), the reported group-level performance differences may partly reflect label uncertainty rather than true model degradation, and should not be over-interpreted.
- Editorial inference: the strong volume correlation suggests a testable extension — using PanSegNet volume trajectories as a quantitative outcome in clinical studies of acute recurrent and chronic pancreatitis, where pancreas volume loss is a candidate biomarker.
- Editorial inference: the four complete failures (one AP, three CP) all involved severe anatomical distortion or atypical anatomy, so collecting more extreme pathological variants, not more healthy controls, is likely the highest-yield next step for improving the model.
- Editorial inference: a natural validation step the paper does not test is human-in-the-loop use — measuring whether showing PanSegNet segmentations to radiologists changes or improves their volume reads.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper evaluates PanSegNet, a deep learning pancreas segmentation model trained on adult MRI/CT data, on a retrospectively collected external pediatric cohort of 84 T2-weighted MRI scans (42 healthy, 23 acute pancreatitis, 19 chronic pancreatitis) at a single center. The authors report Dice Similarity Coefficients of 0.88 (healthy), 0.81 (AP), and 0.80 (CP), with corresponding HD95 values of 3.98 mm, 9.85 mm, and 15.67 mm. They further report volume agreement (R² = 0.85 for healthy, 0.77 for diseased after excluding five cases) and inter/intra-observer agreement metrics. The conclusion claims expert-level performance and positions the work as the first validated deep learning solution for pancreatic MRI segmentation. The paper also shares the annotated dataset and user-interface software.
Significance. If the central claims hold, the paper contributes a valuable external validation of an adult-trained segmentation model in a pediatric population, which is a methodologically strong design because no pediatric labels were used in training. The public release of the dataset and software is a reproducible-resource strength, and the use of standard metrics (DSC, HD95, Jaccard) allows comparison with other studies. However, the strength of the claims of 'expert-level' performance and 'clinical reliability' is limited by the reference-standard noise (inter-observer DSC as low as 0.82), lack of confidence intervals or significance tests, and failure-inclusive volume analysis. The work is a useful contribution, but the evidence as presented does not fully support the strongest claims.
major comments (5)
- [Section 3.3, Table 4] The claim of expert-level performance is not supported by the reported agreement statistics. Inter-observer DSC is 0.82 for pancreatitis and 0.86 for healthy controls, while PanSegNet achieves 0.80-0.81 for AP/CP and 0.88 for healthy controls. The model's performance is therefore within the measured human label noise, but the paper does not report confidence intervals, paired tests, or a non-inferiority analysis against the inter-observer baseline. A bootstrap or paired test would be needed to establish that the model is not worse than human readers, especially in CP where HD95 (15.67 mm) is about double the inter-observer HD95 (7.89 mm).
- [Section 3.2, Figure 4] The volume R² of 0.77 in the diseased group is computed after excluding five cases, and the text separately reports four complete segmentation failures (one AP, three CP). The abstract and conclusion state 'clinical reliability' without a failure-inclusive analysis. The manuscript should report the number of scans used in the regression, the number and reasons for exclusions, and a sensitivity analysis that includes failure cases (e.g., assigning failed segmentations a volume of zero or the reference volume). Without this, the reliability claim applies only to an easier subset of the cohort, which may omit the clinically severe cases (pseudocysts, atrophy, fluid collections).
- [Table 3] The reported standard deviations for SynergyNet are DSC (SD) = 0.80 (1.02) and Jaccard (SD) = 0.70 (1.1). Both exceed the theoretical bounds of these metrics, which are scaled to [0,1]. This appears to be a typographical error (perhaps 0.80 ± 0.02 and 0.70 ± 0.1), but as printed it is a data-integrity issue that undermines the benchmarking table. The authors should correct the values or clarify the notation.
- [Section 2.4] The benchmarking section states that TransUNet, nnUNet3D, nnUNet2D, and SynergyNet were 'trained on our adult pancreas MRI dataset in the same manner as PanSegNet,' but it is not clear whether these models were trained from scratch with the same data, preprocessing, hyperparameters, and training budget, or whether they were fine-tuned from pre-existing weights. The claim that PanSegNet 'outperformed all benchmarked models' requires a controlled training setup; as written, the comparison conditions are ambiguous.
- [Section 3.1, Tables 2 and 3] No confidence intervals or statistical tests are provided for the differences between groups (healthy vs AP vs CP) or between models. The conclusions that PanSegNet achieves 'higher DSC' and 'lower but acceptable HD95' rest on point estimates that may overlap substantially given the reported standard deviations. Adding non-parametric paired comparisons (e.g., Wilcoxon signed-rank tests) and effect-size confidence intervals would make the claims statistically grounded.
minor comments (6)
- [Section 2.1.2 and Table 4] The paper mentions Cohen's kappa for inter- and intra-observer agreement but does not define how kappa is computed for volumetric segmentations (e.g., voxel-level binary classification). Please specify the exact computation, since kappa is not standard for overlap metrics like DSC.
- [Table 1] The percentage in the pancreatitis column uses a comma as a decimal separator ('54,76%'). Use a period for consistency with the rest of the text.
- [Table 2] The 'Accuracy (SD)' column (0.99 ± 0.0) is not defined. If it is voxel-wise accuracy, it is uninformative for a small-organ segmentation problem due to class imbalance and should be removed or replaced with a meaningful metric.
- [Discussion, first paragraph] The discussion states PanSegNet achieved 'DSC of 0.88 in healthy subjects and 0.85 in the pancreatitis group,' but Table 2 reports 0.80 (CP) and 0.81 (AP). Clarify whether the 0.85 is a pooled or weighted average, and reconcile the discrepancy.
- [Supplementary, Figure S8b] The caption 'Belonged the subsequent slice to the same patients; not consistent with image acquisition protocol' is grammatically unclear and should be rewritten.
- [Introduction, firsts list] The bullet list of 'firsts' is promotional and contains a strong claim ('first validated deep learning solution') that is not fully supported by the single-center retrospective design. Consider tempering the wording.
Circularity Check
External pediatric validation of an adult-trained model; self-citation of the model is not load-bearing.
full rationale
The paper's central claim is that PanSegNet, trained on an adult MRI dataset in the authors' prior work (ref 12), transfers to pediatric T2-weighted MRI without retraining. The pediatric cohort (84 scans from Gazi University) is entirely external to the training set, and the paper explicitly states that no internal data splitting was performed and the pediatric data were used only for external testing and validation. No equation in the paper defines the pediatric DSC/HD95 values in terms of the training data; these are measured outcomes on an independent cohort. The volume R^2 values are standard regression summaries of agreement between automated and manual volumes, not predictions derived from fitted parameters. The reference standard is manual segmentation, and the paper reports inter-observer variability; that is a measurement-quality concern, not a circularity. The only self-citation is the PanSegNet model and adult training data from ref 12, which is necessary provenance for the transfer-learning claim but does not by itself establish the pediatric results. Benchmarking models were also trained on the authors' adult dataset, which is a fairness limitation but not circularity. Overall, the derivation chain is self-contained with respect to the pediatric evaluation; no step reduces to its own input.
Assumptions & free parameters
assumptions (4)
- domain assumption Manual radiologist segmentations are treated as the ground truth reference standard.
- domain assumption A model trained on adult pancreas MRI transfers to pediatric pancreas MRI without adaptation.
- domain assumption T2-weighted MRI is the appropriate sequence for pancreas segmentation in this cohort.
- domain assumption Excluded low-quality and failed cases, if included, would not overturn the reported volume agreement.
Cite this review
Pith. "Pith review of Pediatric Pancreas Segmentation from MRI Scans with Deep Learning." pith.science (2026). https://pith.science/paper/EBZ6R3U4
@misc{pith2026250615908,
author = {Pith},
title = {Pith review of: Pediatric Pancreas Segmentation from MRI Scans with Deep Learning},
year = {2026},
howpublished = {\url{https://pith.science/paper/EBZ6R3U4}},
note = {Machine review of arXiv:2506.15908}
}
read the original abstract
Objective: Our study aimed to evaluate and validate PanSegNet, a deep learning (DL) algorithm for pediatric pancreas segmentation on MRI in children with acute pancreatitis (AP), chronic pancreatitis (CP), and healthy controls. Methods: With IRB approval, we retrospectively collected 84 MRI scans (1.5T/3T Siemens Aera/Verio) from children aged 2-19 years at Gazi University (2015-2024). The dataset includes healthy children as well as patients diagnosed with AP or CP based on clinical criteria. Pediatric and general radiologists manually segmented the pancreas, then confirmed by a senior pediatric radiologist. PanSegNet-generated segmentations were assessed using Dice Similarity Coefficient (DSC) and 95th percentile Hausdorff distance (HD95). Cohen's kappa measured observer agreement. Results: Pancreas MRI T2W scans were obtained from 42 children with AP/CP (mean age: 11.73 +/- 3.9 years) and 42 healthy children (mean age: 11.19 +/- 4.88 years). PanSegNet achieved DSC scores of 88% (controls), 81% (AP), and 80% (CP), with HD95 values of 3.98 mm (controls), 9.85 mm (AP), and 15.67 mm (CP). Inter-observer kappa was 0.86 (controls), 0.82 (pancreatitis), and intra-observer agreement reached 0.88 and 0.81. Strong agreement was observed between automated and manual volumes (R^2 = 0.85 in controls, 0.77 in diseased), demonstrating clinical reliability. Conclusion: PanSegNet represents the first validated deep learning solution for pancreatic MRI segmentation, achieving expert-level performance across healthy and diseased states. This tool, algorithm, along with our annotated dataset, are freely available on GitHub and OSF, advancing accessible, radiation-free pediatric pancreatic imaging and fostering collaborative research in this underserved domain.
Reference graph
Works this paper leans on
-
[1]
Pancreatitis is an emerging inflammatory pancreatic disorder of childhood (1, 2)
INTRODUCTION Pancreatitis in Childhood: A Growing Public Health Concern. Pancreatitis is an emerging inflammatory pancreatic disorder of childhood (1, 2). The incidence of pediatric AP has increased over the last ~20 years, affecting an estimated ~ 13 cases per 100,000 children per year (1, 3). Acute pancreatitis (AP) is characterized by sudden inflammati...
-
[2]
METHODS 2.1 Dataset 2.1.1 Data collection and inclusion/exclusion criteria With IRB approval, we retrospectively collected 84 MRI scans (either 1.5T or 3T from Siemens Aera and Verio Scanners) from healthy children who underwent MRI due to the detected hypoechoic liver lesion on abdominal sonography. Additionally, children with AP or CP were diagnosed acc...
work page 2015
-
[4]
DISCUSSION Key findings and clinical value. Accurate segmentation of the pancreas in pediatric patients is crucial for diagnosing and managing pancreatic diseases. This study evaluated PanSegNet, our effective DL approach for pediatric pancreas segmentation, using T2W MRI scans from patients with AP, CP, and healthy subjects. PanSegNet achieved a DSC of 0...
work page 2019
-
[10]
Deep learning algorithms for pancreas segmentation from radiology scans: A review
Zhang Z, Yao L, Keles E, Velichko Y , Bagci U. Deep learning algorithms for pancreas segmentation from radiology scans: A review. Advances in Clinical Radiology. 2023;5(1):31-52. 11. Keles E, Bagci U. The past, current, and future of neonatal intensive care units with artificial intelligence: a systematic review. npj Digital Medicine. 2023;6(1):220. 12. Zh...
work page 2023
-
[26]
Moore MM, Iyer RS, Sarwani NI, Sze RW. Artificial intelligence development in pediatric body magnetic resonance imaging: best ideas to adapt from adults. Pediatric Radiology. 2021;52(2):367-73. 27. Adegun AA, Viriri S, Ogundokun RO. Deep Learning Approach for Medical Image Analysis. Computational Intelligence and Neuroscience. 2021;2021(1):6215281. 28. Dha...
work page 2021
-
[42]
Gorade V , Mittal S, Jha D, Bagci U, editors. Synergynet: Bridging the gap between discrete and continuous representations for precise medical image segmentation. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision; 2024. 43. Bagci U, Chen X, Udupa JK. Hierarchical scale-based multiobject recognition of 3-D anatomical structur...
work page 2024
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.