Pith. sign in

REVIEW 2 major objections 4 minor 25 references

Cancer-Net PCa-Seg: Benchmarking Deep Learning Models for Prostate Cancer Segmentation Using Synthetic Correlated Diffusion Imaging

T0 review · 2 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read SegResNet wins prostate gland segmentation benchmark on CDI^s MRI data.

desk verdict A sound but modest benchmark of five segmentation models on CDI^s prostate gland segmentation; the 'superior' claim for SegResNet needs statistical backing. read the letter →

arxiv 2501.09185 v2 pith:NOHYSEI3 submitted 2025-01-15 eess.IV cs.CV

classification eess.IVcs.CV
keywords prostatecancerglandsegmentationsyntheticcorrelateddiffusionimagingdeeplearningDice-Sørensencoefficientmedicalimagebenchmarkmagneticresonance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether deep learning models can segment prostate glands accurately on synthetic correlated diffusion imaging (CDI^s), a newer MRI modality that sharpens contrast between cancerous and healthy tissue. Using the 200-patient Cancer-Net PCa-Data cohort, the authors benchmark five architectures — U-Net, SegResNet, Swin UNETR, Attention U-Net, and LightM-UNet — and report that SegResNet achieves the best test-set Dice-Sørensen coefficient of 76.68 ± 0.8. Attention U-Net comes within about two points (74.82 ± 2.0) while using roughly a third of SegResNet's parameters, which the authors present as the better accuracy–efficiency tradeoff. The point of the benchmark is to guide model selection for prostate cancer management, where accurate gland delineation is a prerequisite for lesion localization.

What carries the argument

The central object is synthetic correlated diffusion imaging (CDI^s), an MRI-derived modality that the paper uses because it enhances the contrast between clinically relevant and healthy tissue. The argument is carried by the Cancer-Net PCa-Data benchmark, a 200-patient CDI^s dataset with gland and lesion masks derived from PROSTATEx, and by a standardized evaluation protocol: patient-level 70/15/15 splits, five random fold seeds, 200 training epochs from scratch, and reporting of Dice-Sørensen coefficient, inference time, and parameter count. These components make the model ranking interpretable and reproducible.

What would settle it

Retrain the same five models on an independent CDI^s cohort whose gland masks have been re-annotated by expert radiologists, then compare test-set Dice rankings with confidence intervals; SegResNet's claimed superiority would be falsified if another model consistently achieves equal or better Dice, or if the observed gap shrinks below statistical significance.

Watch

Extended reading notes

Core claim

On the Cancer-Net PCa-Data test sets, SegResNet attained the highest Dice-Sørensen coefficient for prostate gland segmentation, 76.68 ± 0.8, outperforming U-Net (71.35 ± 1.5), Swin UNETR (73.46 ± 1.8), Attention U-Net (74.82 ± 2.0), and LightM-UNet (71.38 ± 1.5) across five seeded folds. The authors conclude that SegResNet is the most accurate of the evaluated models for delineating prostate gland boundaries on CDI^s, but note that its higher parameter count and inference time make it less suitable for high-throughput clinical use. They single out Attention U-Net as a nearly as accurate and considerably lighter alternative. The results are presented as evidence that CDI^s, combined with standard deep learning segmentation models, can support prostate cancer management by providing reliable anatomical context.

Load-bearing premise

The ranking stands only if the 200-patient Cancer-Net PCa-Data cohort and its PROSTATEx-derived gland masks are representative of the broader prostate cancer population and its ground truth; if they are biased, the reported Dice scores and the model ordering may not transfer to clinical practice.

Editorial extensions

If this is right

  • SegResNet should be the default architecture when prostate gland segmentation accuracy on CDI^s is the top priority.
  • Attention U-Net is the practical choice for clinical deployment where memory or throughput matters, since it sacrifices only about 1.9 Dice points for roughly 68% fewer parameters.
  • Transformer and Mamba variants (Swin UNETR, LightM-UNet) did not beat well-tuned CNNs on this dataset, suggesting that architecture novelty alone is not decisive for CDI^s gland segmentation.
  • The gland segmentation DSCs reported here (71–77) are far above prior lesion-level DSCs (roughly 0.28–0.32) on mpMRI, indicating that gland-level segmentation on CDI^s is an achievable first stage in a PCa pipeline.
  • The patient-level split and multi-seed protocol provide a robust evaluation baseline that future CDI^s segmentation studies can directly compare against.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural next step the paper does not test is chaining SegResNet's gland masks into a lesion-detection model: if CDI^s really sharpens tumor contrast, constraining detection to the gland could cut false positives.
  • Because all models were trained from scratch with the same budget, the transformer and Mamba architectures may be under-credited; pre-trained encoders or longer schedules could plausibly change the ranking.
  • The efficiency profile of Attention U-Net suggests an even lighter distilled model could approach SegResNet's accuracy, which would be worth testing in resource-limited settings.
  • A direct comparison on the same patients between CDI^s and T2-weighted or mpMRI inputs would isolate whether the modality itself, rather than the model, drives the reported performance.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper benchmarks five deep learning segmentation architectures (U-Net, SegResNet, Swin UNETR, Attention U-Net, and LightM-UNet) on synthetic correlated diffusion imaging (CDI^s) volumes from the Cancer-Net PCa-Data cohort of 200 prostate cancer patients. The authors use a patient-level 70/15/15 train/validation/test split, train each model for 200 epochs with five different data fold seeds, and report Dice-Sørensen coefficient (DSC), inference time, and parameter count. They report that SegResNet achieves the highest mean test DSC (76.68 ± 0.8) and that Attention U-Net offers a favorable accuracy-efficiency tradeoff, concluding that SegResNet is superior for prostate cancer segmentation on this dataset.

Significance. The study provides a useful empirical benchmark for a newly proposed synthetic MRI modality (CDI^s) and uses a sound evaluation protocol: patient-level splitting avoids data leakage, and five seeds give some sense of fold variability. The public dataset and the use of standard MONAI implementations are strengths for reproducibility. However, the headline 'superior' claim for SegResNet is not statistically substantiated, and the manuscript's inconsistent use of 'gland' versus 'tumor' segmentation obscures the clinical scope. As a descriptive comparison of model architectures on a single dataset, the paper is informative, but as a claim of SegResNet superiority it requires revision. The stress-test concern about missing significance testing is valid and is the primary basis for my recommendation.

major comments (2)
  1. [Section 3, Table 1] The claim that SegResNet achieved 'superior prostate tumor segmentation capabilities' is not statistically supported. With n=5 folds, the 1.86-point DSC gap between SegResNet (76.68 ± 0.8) and Attention U-Net (74.82 ± 2.0) is within the Attention U-Net's reported fold-level standard deviation, and no paired significance test (e.g., Wilcoxon signed-rank test) or confidence intervals are provided. The observed ordering could be due to seed selection, so the word 'superior' should be replaced with a descriptive statement about the highest mean DSC, or the authors should add a proper significance test on the per-fold DSCs.
  2. [Section 3 and Abstract] The manuscript conflates prostate gland segmentation with prostate tumor/lesion segmentation. Section 2.1 describes the dataset as containing both prostate gland and tumor lesion masks, and Figure 2 and the text indicate that the evaluated masks are whole-gland masks. The Results sentence 'SegResNet achieved the highest test set DSC... indicating superior prostate tumor segmentation capabilities' is therefore inaccurate if the mask is the whole gland; it should say 'prostate gland segmentation.' This terminology appears in the abstract, the Results, and Figure 1's caption, and it affects the clinical interpretation of the reported results.
minor comments (4)
  1. [Section 2.3] The description of the StepLR scheduler is ambiguous: 'a step size of 0.2 adjusting the learning rate every 50 epochs' is not a standard configuration, since StepLR step_size is normally an integer number of epochs and a gamma decay factor should be specified. Please clarify the exact scheduler parameters, as this directly affects reproducibility.
  2. [Section 1] The introduction states that deep learning methods achieve 'statistically comparable performance metrics to those of experienced radiologists' and cites reference [6], but the details of that statistical comparison are not provided; consider clarifying which study and endpoints are being referenced.
  3. [Section 3] Only aggregate mean and standard deviation across the five seeds are reported for DSC; providing a per-fold table or a supplementary figure would allow readers to assess the distribution of results and the effect of seed selection.
  4. [Abstract and Title] The title and abstract use the phrase 'prostate cancer segmentation,' which could imply lesion or tumor segmentation; since the evaluated task is gland segmentation, the title and abstract should be adjusted to explicitly say 'prostate gland segmentation' for consistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark results are independent empirical measurements on held-out data.

full rationale

The paper's central claim, that SegResNet achieved the highest test-set DSC on the Cancer-Net PCa-Data dataset, is an empirical benchmark result rather than a quantity derived from the dataset's construction, the model definitions, or any fitted equation. The dataset is introduced in prior work by the same group (Ref [12]) and the synthetic correlated diffusion imaging modality originates from Ref [9] by overlapping authors, but these are inputs to the experiment, not outputs of the derivation chain. DSCs reported in Table 1 are measured on a held-out test split (15% of patients) across five fold seeds, and no equation in the paper re-expresses a fitted parameter as a prediction. The only substantive weaknesses are statistical and semantic: the 'superior' claim rests on a 1.86 DSC gap with n=5 folds and no significance test, and the text says 'prostate tumor segmentation' while the evaluation masks are whole-gland masks. These concerns bear on correctness and generalization, not on circularity. The self-citations supply the data and imaging modality, but the benchmark itself is an independent measurement that could in principle falsify or support prior claims made by the same group. Therefore no circular step is exhibited by the paper's own equations or construction, and the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the accuracy and representativeness of the dataset and on the chosen training hyperparameters, none of which are given independent justification. There are no invented physical entities or derived formulas; the study is empirical, so the free parameters are training choices rather than fitted model parameters.

free parameters (4)
  • Learning rate = 0.001 (0.0001 for LightM-UNet)
    Chosen per model without a search; affects training dynamics and could alter the final ranking.
  • Batch size = 16
    Fixed for all models; chosen by hand.
  • Input volume size = 128x128x19
    Resizing choice; may affect the fidelity of small gland boundary features.
  • StepLR decay factor = 0.2 every 50 epochs
    Scheduler setting chosen by hand; varying it could change convergence and final performance.
assumptions (4)
  • domain assumption The PROSTATEx-derived gland masks in Cancer-Net PCa-Data are accepted as accurate ground truth.
    Section 2.1 uses the dataset masks without any quality control or secondary verification, so the entire DSC evaluation depends on this assumption.
  • domain assumption CDI^s images preserve sufficient anatomical information to segment the prostate gland.
    No comparison with the source diffusion-weighted or T2-weighted images is provided, so it is assumed that the derived modality retains the needed boundary structure.
  • domain assumption The five 15% test splits (approximately 30 patients each) are representative of the broader prostate cancer population.
    Generalization to clinical use relies on this representativeness; the paper itself notes transferability to other cohorts as future work.
  • domain assumption Min-max normalization and resizing to 128x128x19 do not remove clinically relevant gland boundary information.
    No ablation or alternative preprocessing is tested, so the chosen preprocessing is assumed harmless.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cancer-Net PCa-Seg: Benchmarking Deep Learning Models for Prostate Cancer Segmentation Using Synthetic Correlated Diffusion Imaging." pith.science (2026). https://pith.science/paper/NOHYSEI3

@misc{pith2026250109185,
  author       = {Pith},
  title        = {Pith review of: Cancer-Net PCa-Seg: Benchmarking Deep Learning Models for Prostate Cancer Segmentation Using Synthetic Correlated Diffusion Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NOHYSEI3}},
  note         = {Machine review of arXiv:2501.09185}
}
abstract

Prostate cancer (PCa) is the most prevalent cancer among men in the United States, accounting for nearly 300,000 cases, 29\% of all diagnoses and 35,000 total deaths in 2024. Traditional screening methods such as prostate-specific antigen (PSA) testing and magnetic resonance imaging (MRI) have been pivotal in diagnosis, but have faced limitations in specificity and generalizability. In this paper, we explore the potential of enhancing PCa gland segmentation using a novel MRI modality called synthetic correlated diffusion imaging (CDI$^s$). We employ several state-of-the-art deep learning models, including U-Net, SegResNet, Swin UNETR, Attention U-Net, and LightM-UNet, to segment prostate glands from a 200 CDI$^s$ patient cohort. We find that SegResNet achieved superior segmentation performance with a Dice-Sorensen coefficient (DSC) of $76.68 \pm 0.8$. Notably, the Attention U-Net, while slightly less accurate (DSC $74.82 \pm 2.0$), offered a favorable balance between accuracy and computational efficiency. Our findings demonstrate the potential of deep learning models in improving prostate gland segmentation using CDI$^s$ to enhance PCa management and clinical support.

Figures

Figures reproduced from arXiv: 2501.09185 by the authors.

Figure 1
Figure 1. Visualization of sample data from the patient cohort in Cancer-Net [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Prostate cancer segmentation pipeline using Cancer-Net PCa-Data. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 20 canonical work pages

  1. [1]

    Cancer statistics,

    Rebecca L Siegel, Angela N Giaquinto, and Ahmedin Jemal. Cancer statistics,

  2. [2]

    Prostate cancer facts

    American Cancer Society. Prostate cancer facts. https://www.cancer.org/ cancer/types/prostate-cancer/about/key-statistics.html , 2024. Accessed: 2024-08-30

  3. [3]

    What is the psa test? https://www.cancer

    National Cancer Institute. What is the psa test? https://www.cancer. gov/types/prostate/psa-fact-sheet#: ~:text=the%20PSA%20test%3F-,What% 20is%20the%20PSA%20test%3F,to%20a%20laboratory%20for%20analysis.,

  4. [4]

    Overdiagnosis and overtreatment of prostate cancer

    Stacy Loeb, Marc A Bjurlin, Joseph Nicholson, Teuvo L Tammela, David F Pen- son, H Ballentine Carter, Peter Carroll, and Ruth Etzioni. Overdiagnosis and overtreatment of prostate cancer. European urology, 65(6):1046–1055, 2014

  5. [5]

    Accessed: 2024-08-30

  6. [6]

    Fully automated deep learning model to detect clinically significant prostate cancer at mri

    Jason C Cai, Hirotsugu Nakai, Shiba Kuanar, Adam T Froemming, Candice W Bolan, Akira Kawashima, Hiroaki Takahashi, Lance A Mynderse, Chandler D Dora, Mitchell R Humphreys, et al. Fully automated deep learning model to detect clinically significant prostate cancer at mri. Radiology, 312(2):e232635, 2024

  7. [7]

    Multiparametric mri for prostate cancer diagnosis: current status and future directions

    Armando Stabile, Francesco Giganti, Andrew B Rosenkrantz, Samir S Taneja, Geert Villeirs, Inderbir S Gill, Clare Allen, Mark Emberton, Caroline M Moore, and Veeru Kasivisvanathan. Multiparametric mri for prostate cancer diagnosis: current status and future directions. Nature reviews urology, 17(1):41–61, 2020

  8. [8]

    Deep learning for fully automatic detection, segmentation, and gleason grade estimation of prostate cancer in multiparametric magnetic reso- nance images

    Oscar J Pellicer-Valero, Jose L Marenco Jimenez, Victor Gonzalez-Perez, Juan Luis Casanova Ramon-Borja, Isabel Mart ´ ın Garc ´ ıa, Maria Barrios Benito, Paula Pelechano Gomez, Jos´ e Rubio-Briones, Mar ´ ıa Jos´ e Rup´ erez, and Jos´ e D Mart ´ ın-Guerrero. Deep learning for fully automatic detection, segmentation, and gleason grade estimation of prostat...

Show all 25 references
  1. [9]

    Explainable ai for cnn-based prostate tumor segmentation in multi- parametric mri correlated to whole mount histopathology

    Deepa Darshini Gunashekar, Lars Bielak, Leonard H¨ agele, Benedict Oerther, Matthias Benndorf, Anca-L Grosu, Thomas Brox, Constantinos Zamboglou, and Michael Bock. Explainable ai for cnn-based prostate tumor segmentation in multi- parametric mri correlated to whole mount histo...

  2. [10]

    Enhancing clinical support for breast cancer with deep learning models using synthetic correlated diffusion imaging

    Chi-en Amy Tai, Hayden Gunraj, Nedim Hodzic, Nic Flanagan, Ali Sabri, and Alexander Wong. Enhancing clinical support for breast cancer with deep learning models using synthetic correlated diffusion imaging. In International Workshop on Applications of Medical AI , pages 83–93....

  3. [11]

    Syn- thetic correlated diffusion imaging hyperintensity delineates clinically significant prostate cancer

    Alexander Wong, Hayden Gunraj, Vignesh Sivan, and Masoom A Haider. Syn- thetic correlated diffusion imaging hyperintensity delineates clinically significant prostate cancer. Scientific Reports, 12(1):3376, 2022

  4. [12]

    Cancer-net pca-data: An open-source benchmark dataset for prostate cancer clinical decision support using synthetic correlated diffusion imaging data

    Hayden Gunraj, Chi en Amy Tai, and Alexander Wong. Cancer-net pca-data: An open-source benchmark dataset for prostate cancer clinical decision support using synthetic correlated diffusion imaging data. NeurIPS Workshops , 2023

  5. [13]

    A method of establishing groups of equal amplitude in plant sociology based on similarity of species content and its application to analyses of the vegetation on danish commons

    Thorvald Sorensen. A method of establishing groups of equal amplitude in plant sociology based on similarity of species content and its application to analyses of the vegetation on danish commons. Biologiske skrifter, 5:1–34, 1948. 7

  6. [14]

    Prostatex challenge data [data set], 2017

    Geert Litjens, Oscar Debats, Jelle Barentsz, Nico Karssemeijer, and Henkjan Huisman. Prostatex challenge data [data set], 2017

  7. [15]

    Computer-aided detection of prostate cancer in mri.IEEE Transactions on Medical Imaging , 33(5):1083–1092, 2014

    Geert Litjens, Oscar Debats, Jelle Barentsz, Nico Karssemeijer, and Henkjan Huisman. Computer-aided detection of prostate cancer in mri.IEEE Transactions on Medical Imaging , 33(5):1083–1092, 2014

  8. [16]

    The cancer imaging archive (tcia): Maintain- ing and operating a public information repository

    Kenneth Clark, Bruce Vendt, Kirk Smith, John Freymann, Justin Kirby, Paul Koppel, Stephen Moore, Stanley Phillips, David Maffitt, Michael Pringle, Lawrence Tarbox, and Fred Prior. The cancer imaging archive (tcia): Maintain- ing and operating a public information repository. J...

  9. [17]

    Quality control and whole-gland, zonal and lesion anno- tations for the prostatex challenge public dataset

    Renato Cuocolo, Arnaldo Stanzione, Anna Castaldo, Davide Raffaele De Lucia, and Massimo Imbriaco. Quality control and whole-gland, zonal and lesion anno- tations for the prostatex challenge public dataset. European Journal of Radiology, 138:109647, 2021

  10. [18]

    3d mri brain tumor segmentation using autoencoder regular- ization

    Andriy Myronenko. 3d mri brain tumor segmentation using autoencoder regular- ization. In Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 4th International Workshop, BrainLes 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 16,...

  11. [19]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part ...

  12. [20]

    Attention u-net: Learning where to look for the pancreas

    Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bern- hard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999 , 2018

  13. [21]

    Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images

    Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In International MICCAI brainlesion workshop , pages 272–284. Springer, 2021

  14. [22]

    Monai: Medical open network for ai

    MONAI. Monai: Medical open network for ai. https://monai.io/, 2024. Ac- cessed: 2024-08-30

  15. [23]

    Lightm-unet: Mamba assists in lightweight unet for medical image segmentation

    Weibin Liao, Yinghao Zhu, Xinyuan Wang, Chengwei Pan, Yasha Wang, and Liantao Ma. Lightm-unet: Mamba assists in lightweight unet for medical image segmentation. arXiv preprint arXiv:2403.05246 , 2024

  16. [25]

    Thop: Pytorch-opcounter

    Ligeng Zhu. Thop: Pytorch-opcounter. https://github.com/Lyken17/ pytorch-OpCounter, 2022. Accessed: 2024-07-21. 8

  17. [2024]

    CA: a cancer journal for clinicians , 74(1), 2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.