pith. sign in

arxiv: 1906.10729 · v1 · pith:JYUFIMQGnew · submitted 2019-06-25 · 🧬 q-bio.QM · cs.CV· cs.LG· eess.IV

CNN-based Survival Model for Pancreatic Ductal Adenocarcinoma in Medical Imaging

Yucheng Zhang , Edrise M. Lobo-Mueller , Paul Karanicolas , Steven Gallinger , Masoom A. Haider , Farzad Khalvati This is my paper

Pith reviewed 2026-05-25 15:38 UTC · model grok-4.3

classification 🧬 q-bio.QM cs.CVcs.LGeess.IV
keywords survival analysispancreatic ductal adenocarcinomaconvolutional neural networkradiomicscomputed tomographyprognosistransfer learning
0
0 comments X

The pith

A convolutional neural network survival model using preoperative CT images outperforms traditional Cox proportional hazards radiomics models by 22% in concordance index for pancreatic ductal adenocarcinoma patients.

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

The paper aims to show that a CNN-based survival model can provide better predictions of patient survival from CT scans than standard Cox proportional hazards models used in radiomics. The CPH model relies on linear assumptions and struggles with correlated features, which the CNN avoids by learning directly from images via transfer learning. Tested on preoperative CT of resectable PDAC patients, the CNN achieved a 22% higher concordance index. This suggests deep learning can better capture the complex relationships in imaging data for prognosis. A sympathetic reader would care because improved survival prediction could help in treatment planning for pancreatic cancer.

Core claim

Using transfer learning, a convolutional neural network based survival model was built and tested on preoperative CT images of resectable Pancreatic Ductal Adenocarcinoma patients. The proposed CNN-based survival model outperformed the traditional CPH-based radiomics approach in terms of concordance index by 22%, providing a better fit for patients' survival patterns. The proposed CNN-based survival model outperforms CPH-based radiomics pipeline in PDAC prognosis. This approach offers a better fit for survival patterns based on CT images and overcomes the limitations of conventional survival models.

What carries the argument

Transfer learning CNN for direct survival ranking from CT images, replacing radiomic feature extraction and CPH modeling.

If this is right

  • CNN provides better fit for survival patterns from CT images.
  • Overcomes linear assumption and multicollinearity issues in CPH models.
  • Improved prognostic performance for PDAC using preoperative imaging.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method could be applied to other cancers with available CT data for survival prediction.
  • Further testing on external validation sets is implied to ensure the model is not overfit.
  • This could lead to revised clinical guidelines for using imaging in survival estimation.

Load-bearing premise

The CNN trained via transfer learning on the available preoperative CT cohort produces a generalizable ranking of survival times without overfitting or dataset-specific artifacts.

What would settle it

Performance evaluation on an independent external validation cohort of PDAC patients from different institutions showing the concordance index improvement does not hold would falsify the superiority claim.

read the original abstract

Cox proportional hazard model (CPH) is commonly used in clinical research for survival analysis. In quantitative medical imaging (radiomics) studies, CPH plays an important role in feature reduction and modeling. However, the underlying linear assumption of CPH model limits the prognostic performance. In addition, the multicollinearity of radiomic features and multiple testing problem further impedes the CPH models performance. In this work, using transfer learning, a convolutional neural network (CNN) based survival model was built and tested on preoperative CT images of resectable Pancreatic Ductal Adenocarcinoma (PDAC) patients. The proposed CNN-based survival model outperformed the traditional CPH-based radiomics approach in terms of concordance index by 22%, providing a better fit for patients' survival patterns. The proposed CNN-based survival model outperforms CPH-based radiomics pipeline in PDAC prognosis. This approach offers a better fit for survival patterns based on CT images and overcomes the limitations of conventional survival models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes a CNN-based survival model trained via transfer learning on preoperative CT images of resectable PDAC patients. It claims this model outperforms a traditional CPH-based radiomics pipeline by 22% in concordance index, addressing the linear assumption, multicollinearity, and multiple-testing limitations of CPH while providing a better fit to survival patterns.

Significance. If the performance gain is shown to be robust under proper held-out validation on a sufficiently large cohort, the work would demonstrate that deep learning can usefully relax the restrictive assumptions of CPH in radiomics survival modeling, offering a practical alternative for PDAC prognosis from imaging data.

major comments (3)
  1. [Abstract] Abstract: the headline claim of a 22% C-index improvement is presented without any report of cohort size (n), cross-validation protocol, confidence intervals, or statistical testing; without these quantities the result cannot be evaluated for stability or generalizability.
  2. [Abstract] Abstract and Methods (implied): the description of how the CNN output is converted into a survival ranking (e.g., whether a Cox partial-likelihood loss, ranking loss, or discrete-time hazard is used) is absent, leaving the precise modeling innovation unspecified.
  3. [Abstract] Abstract: no comparison to non-radiomics baselines (e.g., clinical variables alone or standard image CNNs without radiomics) is supplied, so it is unclear whether the reported gain is attributable to the CNN architecture or simply to the use of imaging features at all.
minor comments (1)
  1. [Abstract] Abstract: the phrase 'providing a better fit for patients' survival patterns' is vague; a quantitative statement of calibration or time-dependent AUC would be more informative.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive comments on our manuscript. We provide point-by-point responses to the major comments below. We will revise the abstract to address the first two points.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline claim of a 22% C-index improvement is presented without any report of cohort size (n), cross-validation protocol, confidence intervals, or statistical testing; without these quantities the result cannot be evaluated for stability or generalizability.

    Authors: We agree with the referee that the abstract should provide these essential details to allow proper evaluation of the results. Although the full manuscript describes the cohort size, cross-validation protocol, and includes confidence intervals and statistical comparisons in the Results section, we will revise the abstract to concisely include this information for improved clarity and to highlight the robustness of the findings. revision: yes

  2. Referee: [Abstract] Abstract and Methods (implied): the description of how the CNN output is converted into a survival ranking (e.g., whether a Cox partial-likelihood loss, ranking loss, or discrete-time hazard is used) is absent, leaving the precise modeling innovation unspecified.

    Authors: We acknowledge that the abstract does not specify the exact survival modeling approach used with the CNN output. We will update the abstract to briefly describe the method by which the CNN produces survival rankings, including the loss function or ranking approach employed. This is elaborated in the Methods section of the manuscript. revision: yes

  3. Referee: [Abstract] Abstract: no comparison to non-radiomics baselines (e.g., clinical variables alone or standard image CNNs without radiomics) is supplied, so it is unclear whether the reported gain is attributable to the CNN architecture or simply to the use of imaging features at all.

    Authors: The comparison presented is between the proposed CNN-based survival model and a traditional CPH-based radiomics pipeline, both of which utilize features derived from the preoperative CT images. This design isolates the effect of the modeling approach (CNN vs. CPH with radiomics) rather than the use of imaging versus non-imaging data. Therefore, the 22% improvement can be attributed to the CNN architecture relaxing the assumptions of CPH. We do not believe additional non-radiomics baselines are necessary to support this specific claim, though we can add a discussion clarifying this point if the editor deems it helpful. revision: no

Circularity Check

0 steps flagged

No circularity: empirical head-to-head comparison on imaging data

full rationale

The paper reports an empirical performance comparison between a transfer-learned CNN survival model and a CPH radiomics baseline on preoperative CT scans for PDAC. The 22% C-index gain is a measured metric on the cohort; no equations, ansatzes, or uniqueness theorems are present that could reduce the reported result to a fitted parameter or self-citation by construction. The central claim rests on data-driven evaluation rather than any self-referential derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; the approach rests on standard transfer learning and the unstated assumption that the CNN architecture can capture non-linear survival relationships from CT voxels.

pith-pipeline@v0.9.0 · 5732 in / 1040 out tokens · 23626 ms · 2026-05-25T15:38:22.692619+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages

  1. [1]

    All convolutional layers have kernel size of 33 with 32 filters following by Batch Normalization layers (BN)

    Input images have dimensions of 1401401 (grey scale), which contain the CT images within the manual contours of the tumors (example shown in Figure 2). All convolutional layers have kernel size of 33 with 32 filters following by Batch Normalization layers (BN). The first Max Pool layer has pool size of 2×2, and the latter two Max Pool layers have pool ...

    work page

  2. [2]

    & Aban, I

    George, B., Seals, S. & Aban, I. Survival analysis and regression models. J. Nucl. Cardiol. 21, 686–94 (2014)

    work page 2014

  3. [3]

    Katzman, J. et al. DeepSurv: Personalized Treatment Recommender System Using A Cox Proportional Hazards Deep Neural Network. (2016). doi:10.1186/s12874-018-0482-1

    work page doi:10.1186/s12874-018-0482-1 2016

  4. [4]

    Gensheimer, M. F. & Narasimhan, B. A Scalable Discrete-Time Survival Model for Neural Networks

    work page

  5. [5]

    & Garmire, L

    Ching, T., Zhu, X. & Garmire, L. X. Cox-nnet: An artificial neural network method for prognosis prediction of high-throughput omics data. PLoS Comput. Biol. 14, e1006076 (2018)

    work page 2018

  6. [6]

    Cox, D. R. Regression Models and Life-Tables. Journal of the Royal Statistical Society. Series B (Methodological) 34, 187–220 (1972)

    work page 1972

  7. [7]

    Katzman, J. L. et al. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Med. Res. Methodol. 18, 24 (2018)

    work page 2018

  8. [8]

    Random Forests

    Breiman, L. Random Forests. 1–33 (2001)

    work page 2001

  9. [9]

    A., Dumais, S

    Hearst, M. A., Dumais, S. T., Osman, E., Platt, J. & Scholkopf, B. Support vector machines. IEEE Intell. Syst. 13, 18–28 (1998)

    work page 1998

  10. [10]

    & Falasca, M

    Adamska, A., Domenichini, A. & Falasca, M. Pancreatic Ductal Adenocarcinoma: Current and Evolving Therapies. Int. J. Mol. Sci. 18, (2017)

    work page 2017

  11. [11]

    Foucher, E. D. et al. Pancreatic Ductal Adenocarcinoma: A Strong Imbalance of Good and Bad Immunological Cops in the Tumor Microenvironment. Front. Immunol. 9, 1044 (2018)

    work page 2018

  12. [12]

    Stark, A. P. et al. Long-term survival in patients with pancreatic ductal adenocarcinoma. Surgery 159, 1520–1527 (2016)

    work page 2016

  13. [13]

    Mariani, L. et al. Prognostic factors for metachronous contralateral breast cancer: a comparison of the linear Cox regression model and its artificial neural network extension. Breast Cancer Res. Treat. 44, 167–78 (1997)

    work page 1997

  14. [14]

    & Azen, S

    Xiang, A., Lapuerta, P., Ryutov, A., Buckley, J. & Azen, S. Comparison of the performance of neural network methods and Cox regression for censored survival data. Comput. Stat. Data Anal. 34, 243–257 (2000)

    work page 2000

  15. [15]

    Kumar, V. et al. Radiomics: the process and the challenges. Magn. Reson. Imaging 30, 1234–1248 (2012)

    work page 2012

  16. [16]

    A., Leijenaar, R

    Keek, S. A., Leijenaar, R. T., Jochems, A. & Woodruff, H. C. A review on radiomics and the future of theranostics for patient selection in precision medicine. Br. J. Radiol. 91, 20170926 (2018)

    work page 2018

  17. [17]

    Zhang, Y., Oikonomou, A., Wong, A., Haider, M. A. & Khalvati, F. Radiomics-based Prognosis Analysis for Non-Small Cell Lung Cancer. Nat. Sci. Reports 7, (2017)

    work page 2017

  18. [18]

    Aerts, H. J. et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun 5, 4006 (2014)

    work page 2014

  19. [19]

    Lambin, P. et al. Radiomics: Extracting more information from medical images using advanced feature analysis. Eur. J. Cancer 48, 441–446 (2012)

    work page 2012

  20. [20]

    Aerts, H. J. W. L. The Potential of Radiomic-Based Phenotyping in Precision Medicine. JAMA Oncol. 2, 1636 (2016)

    work page 2016

  21. [21]

    van Griethuysen, J. J. M. et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res. 77, e104–e107 (2017)

    work page 2017

  22. [22]

    Khalvati, F. et al. Prognostic Value of CT Radiomic Features in Resectable Pancreatic Ductal Adenocarcinoma. Sci. Rep. 9, 5449 (2019)

    work page 2019

  23. [23]

    Huang, Y. et al. Radiomics Signature: A Potential Biomarker for the Prediction of Disease-Free Survival in Early-Stage (I or II) Non-Small Cell Lung Cancer. Radiology 152234 (2016). doi:10.1148/radiol.2016152234

    work page doi:10.1148/radiol.2016152234 2016

  24. [24]

    & Hinton, G

    LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444 (2015)

    work page 2015

  25. [25]

    & Hinton, G

    Krizhevsky, A., Sutskever, I. & Hinton, G. E. ImageNet Classification with Deep Convolutional Neural Networks. 1097–1105 (2012)

    work page 2012

  26. [26]

    Shin, H.-C. et al. Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning. IEEE Trans. Med. Imaging 35, 1285–1298 (2016)

    work page 2016

  27. [27]

    Tajbakhsh, N. et al. Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning? IEEE Trans. Med. Imaging 35, 1299–1312 (2017)

    work page 2017

  28. [28]

    & Kiryu, S

    Yasaka, K., Akai, H., Abe, O. & Kiryu, S. Deep Learning with Convolutional Neural Network for Differentiation of Liver Masses at Dynamic Contrast-enhanced CT: A Preliminary Study. Radiology 286, 887–896 (2018)

    work page 2018

  29. [29]

    Yamashita, R., Nishio, M., Do, R. K. G. & Togashi, K. Convolutional neural networks: an overview and application in radiology. Insights Imaging 9, 611–629 (2018)

    work page 2018

  30. [30]

    Aerts, H. J. W. L. et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat. Commun. 5, 4006 (2014)

    work page 2014

  31. [31]

    C., Steck, H., Krishnapuram, B., Dehing-Oberije, C

    Raykar, V. C., Steck, H., Krishnapuram, B., Dehing-Oberije, C. & Lambin, P. On Ranking in Survival Analysis: Bounds on the Concordance Index

    work page

  32. [32]

    Haibe-Kains, B. et al. survcomp: a package for performance assessment and comparison for survival analysis. (2019)

    work page 2019

  33. [33]

    Schmid, M., Wright, M. N. & Ziegler, A. On the use of Harrell’s C for clinical risk prediction via random survival forests. (2016)

    work page 2016