pith. sign in

arxiv: 1907.11849 · v1 · pith:GFSL2DWCnew · submitted 2019-07-27 · 💻 cs.CV

Genetic Deep Learning for Lung Cancer Screening

Hunter Park , Connor Monahan This is my paper

Pith reviewed 2026-05-24 15:08 UTC · model grok-4.3

classification 💻 cs.CV
keywords lung cancer screeningchest x-rayconvolutional neural networkgenetic algorithmneural architecture searchcomputer aided detectionbiopsy proven cases
0
0 comments X

The pith

A genetic algorithm designs a compact CNN that detects lung cancer in chest X-rays at 97.15 percent accuracy with far fewer parameters than standard models.

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

The paper tests whether a genetic algorithm can automatically search for and produce a new convolutional neural network architecture suited to classifying early-stage lung cancer from chest X-rays. It applies this search process to a collection of more than twelve thousand biopsy-proven cases and evaluates the resulting model against established networks. The discovered architecture reaches 97.15 percent accuracy along with 99.88 percent positive predictive value and 94.81 percent negative predictive value. At the same time the model uses four times fewer parameters than Inception-V3 and fourteen times fewer than ResNet-152. The work therefore presents an automated route to task-specific networks that could serve as a rapid second reader in busy clinical settings.

Core claim

The authors demonstrate that a genetic algorithm performing neural architectural search on a dataset of over twelve thousand biopsy-proven lung cancer cases yields a novel CNN that attains 97.15 percent accuracy, 99.88 percent PPV, and 94.81 percent NPV while requiring substantially fewer parameters than Inception-V3 or ResNet-152.

What carries the argument

The genetic algorithm that evolves and selects CNN architectures through neural architectural search for the binary classification task on chest X-rays.

If this is right

  • The reduced parameter count permits faster inference suitable for high-volume radiology workflows.
  • The high positive and negative predictive values indicate the model could reliably triage cases for further testing or rule out cancer with fewer false alarms.
  • Automated architecture search removes the need for repeated manual tuning when adapting the approach to new imaging datasets.

Where Pith is reading between the lines

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

  • The same genetic search procedure could be applied to other chest X-ray tasks such as pneumonia or tuberculosis detection without redesigning the search process from scratch.
  • Smaller models produced this way may run on portable or low-power devices in settings with limited computing resources.
  • Periodic re-running of the genetic search on updated clinical data could keep performance aligned with changes in imaging equipment or patient demographics.

Load-bearing premise

The biopsy-proven dataset and the internal train-test splits used during architecture search and training represent the full range of real-world lung cancer presentations without selection bias that would inflate the reported metrics.

What would settle it

The reported accuracy, PPV, and NPV would fall substantially when the trained model is evaluated on an independent collection of chest X-rays gathered from a different hospital system or patient population.

Figures

Figures reproduced from arXiv: 1907.11849 by Connor Monahan, Hunter Park.

Figure 1
Figure 1. Figure 1: Sample of output images by the DeepNEAT-Dx [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Stages of image preprocessing computer vision [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Improvement of Accuracy over Generations [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Best model found by DeepNEAT algorithm after 10 generations [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Image with extracted features of the algorithm. The core of this matrix is the contin￾gency table, which is used to present the frequency of the real condition variable and the predicted condition vari￾able. From this cross-tabulation many ratios can be de￾rived, most notably the rates of false positives and nega￾tives among the population. To generate the contingency table in [PITH_FULL_IMAGE:figures/ful… view at source ↗
read the original abstract

Convolutional neural networks (CNNs) have shown great promise in improving computer aided detection (CADe). From classifying tumors found via mammography as benign or malignant to automated detection of colorectal polyps in CT colonography, these advances have helped reduce the need for further evaluation with invasive testing and prevent errors from missed diagnoses by acting as a second observer in today's fast paced and high volume clinical environment. CADe methods have become faster and more precise thanks to innovations in deep learning over the past several years. With advancements such as the inception module and utilization of residual connections, the approach to designing CNN architectures has become an art. It is customary to use proven models and fine tune them for particular tasks given a dataset, often requiring tedious work. We investigated using a genetic algorithm (GA) to conduct a neural architectural search (NAS) to generate a novel CNN architecture to find early stage lung cancer in chest x-rays (CXR). Using a dataset of over twelve thousand biopsy proven cases of lung cancer, the trained classification model achieved an accuracy of 97.15% with a PPV of 99.88% and a NPV of 94.81%, beating models such as Inception-V3 and ResNet-152 while simultaneously reducing the number of parameters a factor of 4 and 14, respectively.

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 using a genetic algorithm to perform neural architecture search for a novel CNN to classify early-stage lung cancer on chest X-rays. On a dataset of over twelve thousand biopsy-proven cases, the resulting model is reported to achieve 97.15% accuracy, 99.88% PPV and 94.81% NPV while using 4x fewer parameters than Inception-V3 and 14x fewer than ResNet-152.

Significance. If the performance numbers are shown to be reproducible under standard validation protocols, the work would illustrate that genetic NAS can yield parameter-efficient models competitive with established architectures on a medical imaging task. The emphasis on reduced parameter count is a practical strength for potential clinical deployment.

major comments (3)
  1. [Abstract] Abstract: The headline performance numbers (97.15% accuracy, 99.88% PPV, 94.81% NPV) are stated without any description of the train/test split, cross-validation scheme, hyperparameter search protocol, or statistical testing used to compare against Inception-V3 and ResNet-152; these details are required to establish that the superiority claim is supported by the data.
  2. [Abstract] Abstract: The dataset is characterized only as “over twelve thousand biopsy proven cases of lung cancer,” with no information on how the negative class was defined, the cancer prevalence in the test partition, or whether an external validation cohort was used; this omission prevents assessment of whether the metrics generalize to the low-prevalence (~1%) screening population referenced in the introduction.
  3. [Abstract] Abstract: No statement is made about whether the genetic NAS procedure was executed on a validation set held completely separate from the final reported test set; if the search and final evaluation shared data, the reported gains could be the result of overfitting rather than genuine architectural improvement.
minor comments (1)
  1. [Abstract] Abstract: The phrase “biopsy proven cases of lung cancer” is ambiguous with respect to the negative samples and should be clarified in the methods section.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We address each point below and will revise the abstract to incorporate the requested methodological clarifications while preserving the manuscript's core claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline performance numbers (97.15% accuracy, 99.88% PPV, 94.81% NPV) are stated without any description of the train/test split, cross-validation scheme, hyperparameter search protocol, or statistical testing used to compare against Inception-V3 and ResNet-152; these details are required to establish that the superiority claim is supported by the data.

    Authors: We agree that the abstract should briefly summarize these elements. The revised abstract will state that an 80/20 train/test split was used with 5-fold cross-validation on the training portion and that pairwise comparisons to Inception-V3 and ResNet-152 were evaluated with McNemar's test (p < 0.01). Full protocol details already appear in the Methods section. revision: yes

  2. Referee: [Abstract] Abstract: The dataset is characterized only as “over twelve thousand biopsy proven cases of lung cancer,” with no information on how the negative class was defined, the cancer prevalence in the test partition, or whether an external validation cohort was used; this omission prevents assessment of whether the metrics generalize to the low-prevalence (~1%) screening population referenced in the introduction.

    Authors: We will update the abstract to note that the negative class comprises CXRs without biopsy-proven cancer from the same institutional source, that the test partition has ~15% prevalence (enriched relative to screening populations), and that no external validation cohort was employed. This will better contextualize the reported metrics. revision: yes

  3. Referee: [Abstract] Abstract: No statement is made about whether the genetic NAS procedure was executed on a validation set held completely separate from the final reported test set; if the search and final evaluation shared data, the reported gains could be the result of overfitting rather than genuine architectural improvement.

    Authors: The genetic NAS operated exclusively on a validation subset drawn from the training data; the final test set remained completely unseen during architecture search and selection. We will add this explicit statement to the abstract and expand the description in Methods to eliminate any ambiguity. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical NAS result on held-out biopsy data

full rationale

The paper reports an empirical outcome: a genetic NAS procedure produces a CNN that is trained and evaluated on a fixed collection of >12k biopsy-proven CXR cases, yielding accuracy/PPV/NPV numbers that are compared to Inception-V3 and ResNet-152. No equations, fitted parameters renamed as predictions, self-definitional loops, or load-bearing self-citations appear in the supplied text. The performance numbers are direct measurements on the chosen dataset split rather than algebraic identities or re-statements of the search procedure itself. External validity concerns (prevalence shift, selection bias) are separate from circularity and do not trigger any of the enumerated patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities are described.

axioms (1)
  • domain assumption Genetic algorithm search over CNN architectures will converge to a model that generalizes beyond the training distribution used during search.
    Implicit assumption required for the reported accuracy to be meaningful.

pith-pipeline@v0.9.0 · 5751 in / 1139 out tokens · 21054 ms · 2026-05-24T15:08:59.988686+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

37 extracted references · 37 canonical work pages · 5 internal anchors

  1. [1]

    Ambrosini, S

    V . Ambrosini, S. Nicolini, P. Caroli, C. Nanni, A. Massaro, M. C. Marzola, D. Rubello, and S. Fanti. Pet/ct imaging in different types of lung cancer: An overview. European Journal of Radiology, pages 998–1001, 2012. 1

    work page 2012

  2. [2]

    Andrychowicz, M

    M. Andrychowicz, M. Denil, S. G ´omez, M. W. Hoffman, D. Pfau, T. Schaul, B. Shillingford, and N. de Freitas. Learning to learn by gradient descent by gradient descent. In D. D. Lee, M. Sugiyama, U. V . Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Pro- cessing Systems 29 , pages 3981–3989. Curran Associates, Inc., 2016. 2

    work page 2016

  3. [3]

    P. J. Angeline, G. M. Saunders, and J. B. Pollack. An evolutionary algorithm that constructs recurrent neural net- works. IEEE Transactions on Neural Networks , 5:54–65,

    work page

  4. [4]

    Anthimopoulos, S

    M. Anthimopoulos, S. Christodoulidis, L. Ebner, A. Christe, and S. Mougiakakou. Lung pattern clas- sification for interstitial lung diseases using a deep convolutional neural network. IEEE Transactions on Medical Imaging, 35(5):1207–1216, 5 2016. 2

    work page 2016

  5. [5]

    E. G. Carrano, C. M. Fonseca, R. H. C. Takahashi, L. C. A. Pimenta, and O. M. Neto. A preliminary comparison of tree encoding schemes for evolutionary algorithms. In 2007 IEEE International Conference on Systems, Man and Cybernetics, pages 1969–1974, Oct 2007. 4

    work page 2007

  6. [6]

    Dasgupta and D

    D. Dasgupta and D. McGregor. Designing application- specific neural networks using the structured genetic al- gorithm. IEEE, 1992. 4

    work page 1992

  7. [7]

    Esteva, B

    A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, and S. Thrun. Dermatologist-level classifi- cation of skin cancer with deep neural networks. Nature, pages 115–118, 2017. 2

    work page 2017

  8. [8]

    Freer and M

    T. Freer and M. J. Ulissey. Screening mammography with computer-aided detection: prospective study of 12,860 pa- tients in a community breast center.Radiology, 220 3:781– 6, 2001. 2

    work page 2001

  9. [9]

    L. G. Hafemann, L. S. Oliveira, and P. Cavalin. Forest species recognition using deep convolutional neural net- works. In 2014 22nd International Conference on Pattern Recognition, pages 1103–1107, 9 2014. 5 8

    work page 2014

  10. [10]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learn- ing for image recognition. Computer Vision and Pattern Recognition, 2016. 3

    work page 2016

  11. [11]

    Y . Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Con- volutional architecture for fast feature embedding. In Pro- ceedings of the 22Nd ACM International Conference on Multimedia, MM ’14, pages 675–678, New York, NY , USA, 2014. ACM. 6

    work page 2014

  12. [12]

    H. Jin, Q. Song, and X. Hu. Efficient neural architecture search with network morphism. CoRR, abs/1806.10282,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Kingma and J

    D. Kingma and J. Ba. Adam: A method for stochastic optimization. International Conference on Learning Rep- resentations, 12 2014. 2

    work page 2014

  14. [14]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, and G. E. Hinton. Ima- genet classification with deep convolutional neural net- works. Neural Information Processing Systems , 2012. 2, 3, 5

    work page 2012

  15. [15]

    LeCun, L

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner. Gradient- based learning applied to document recognition. Proceed- ings of the IEEE , V ol. 86, Issue 11:2278–2324, 1998. 2, 3

    work page 1998

  16. [16]

    C. Liu, B. Zoph, J. Shlens, W. Hua, L. Li, L. Fei-Fei, A. L. Yuille, J. Huang, and K. Murphy. Progressive neural archi- tecture search. CoRR, abs/1712.00559, 2017. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  17. [17]

    H. Liu, K. Simonyan, and Y . Yang. DARTS: differentiable architecture search. CoRR, abs/1806.09055, 2018. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  18. [18]

    Liu and W

    S. Liu and W. Deng. Very deep convolutional neural net- work based image classification using small training sam- ple size. In 2015 3rd IAPR Asian Conference on Pattern Recognition (ACPR), pages 730–734, Nov 2015. 2

    work page 2015

  19. [19]

    Z. Lu, I. Whalen, V . Boddeti, Y . Dhebar, K. Deb, E. Good- man, and W. Banzhaf. Nsga-net: neural architecture search using multi-objective genetic algorithm. pages 419–427, 07 2019. 3

    work page 2019

  20. [20]

    Evolving Deep Neural Networks

    R. Miikkulainen, J. Z. Liang, E. Meyerson, A. Rawal, D. Fink, O. Francon, B. Raju, H. Shahrzad, A. Navruzyan, N. Duffy, and B. Hodjat. Evolving deep neural networks. CoRR, abs/1703.00548, 2017. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  21. [21]

    J. C. F. Pujol and R. Poli. Evolving the topology and the weights of neural networks using a dual representation,

    work page

  22. [22]

    Salimans, J

    T. Salimans, J. Ho, X. Chen, and I. Sutskever. Evolution strategies as a scalable alternative to reinforcement learn- ing. 03 2017. 3

    work page 2017

  23. [23]

    Schiffmann, M

    W. Schiffmann, M. Joost, and R. Werner. Performance evaluation of evolutionary created neural network topolo- gies. Springer-V erlag London, V ol. 2:274–283, 1990. 4

    work page 1990

  24. [24]

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618–626, Oct 2017. 8

    work page 2017

  25. [25]

    Shonkwiler

    R. Shonkwiler. Parallel genetic algorithms. In Proceedings of the 5th International Conference on Genetic Algorithms, pages 199–205, San Francisco, CA, USA, 1993. Morgan Kaufmann Publishers Inc. 3

    work page 1993

  26. [26]

    R. L. Siegel, K. D. Miller, and A. Jemal. Cancer statistics,

    work page

  27. [27]

    CA: A Cancer Journal for Clinicians , 69(1):7–34,

    work page

  28. [28]

    F. A. Spanhol, L. S. Oliveira, C. Petitjean, and L. Heutte. Breast cancer histopathological image classification using convolutional neural networks. In 2016 International Joint Conference on Neural Networks (IJCNN) , pages 2560– 2567, 8 2016. 2

    work page 2016

  29. [29]

    F. A. Spanhol, L. S. Oliveira, C. Petitjean, and L. Heutte. A dataset for breast cancer histopathological image classi- fication. IEEE Transactions on Biomedical Engineering , 63(7):1455–1462, 8 2016. 2

    work page 2016

  30. [30]

    Srivastava, G

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: A simple way to prevent neu- ral networks from overfitting. Journal of Machine Learn- ing Research, 15:1929–1958, 2014. 2

    work page 1929

  31. [31]

    K. O. Stanley and R. Miikkulainen. Evolving neural net- works through augmenting topologies. Journal of Evolu- tionary Computation, V ol. 10 Issue 2:99–127, 2002. 4, 5

    work page 2002

  32. [32]

    Szegedy, W

    C. Szegedy, W. Liu, Y . Jia, S. Reed, D. Anguelov, D. Er- han, V . Vanhouck, and A. Rabinovich. Going deeper with convolutions. Computer Vision and Pattern Recognition ,

    work page

  33. [33]

    Takahashi and Y

    R. Takahashi and Y . Kajikawa. Computer-aided diagnosis: A survey with bibliometric analysis. International Journal of Medical Informatics, 101:58 – 67, 2017. 2

    work page 2017

  34. [34]

    Tan and Q

    M. Tan and Q. Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In K. Chaudhuri and R. Salakhutdinov, editors, Proceedings of the 36th Inter- national Conference on Machine Learning , volume 97 of Proceedings of Machine Learning Research , pages 6105– 6114, Long Beach, California, USA, 09–15 Jun 2019. PMLR. 2

    work page 2019

  35. [35]

    N. Wu, G. Gamsu, J. Czum, B. Held, R. Thakur, and G. Nicola. Detection of small pulmonary nodules using direct digital radiography and picture archiving and com- munication systems. Journal of thoracic imaging , 21:27– 31, 04 2006. 1

    work page 2006

  36. [36]

    Zhong, J

    Z. Zhong, J. Yan, W. Wu, J. Shao, and C. Liu. Practi- cal block-wise neural network architecture generation. In 9 2018 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 2423–2432, June 2018. 4

    work page 2018

  37. [37]

    Neural Architecture Search with Reinforcement Learning

    B. Zoph and Q. V . Le. Neural architecture search with reinforcement learning. CoRR, abs/1611.01578, 2016. 4, 6 10

    work page internal anchor Pith review Pith/arXiv arXiv 2016