Pith. sign in

REVIEW 4 major objections 5 minor 17 references

Patch-based Automatic Rosacea Detection Using the ResNet Deep Learning Framework

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read Localized facial patches match or beat full-image inputs for automatic rosacea detection with a ResNet-18.

desk verdict A sensible, clearly described patch-comparison study for rosacea whose central "superior" claim is not yet supported by the reported numbers; worth refereeing, but major revision is needed before it can be believed. read the letter →

arxiv 2509.09841 v1 pith:5FACXP62 submitted 2025-09-11 cs.CV

classification cs.CV
keywords rosaceadetectionpatch-basedclassificationResNet-18regionofinterestprivacy-preservingmedicalimagingdeeplearningdermatologyfacealignment
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 argues that, for automated rosacea detection, feeding a ResNet-18 with localized facial patches—forehead, nose, and cheeks—can be as accurate as, and sometimes more accurate than, feeding it the full face. The authors systematically vary which regions are included and at what scale, and they evaluate the resulting classifiers on a real test set after training on synthetic images. Their results show that multi-region combinations such as cheeks-plus-forehead reach AUC near 1.0 on the real test set, while full-face input sometimes scores lower, and cheeks alone outperform forehead or nose alone. Because only partial facial regions are used, the approach also excludes identifiable facial features, addressing privacy concerns. The paper positions this as a practical trade-off between diagnostic performance, interpretability, and patient confidentiality.

What carries the argument

The mechanism is patch masking of four facial regions of interest (forehead, nose, left cheek, right cheek) at three sizes, applied after eye-coordinate alignment of each face, with a fine-tuned ResNet-18 (ImageNet-pretrained) and a single sigmoid output as classifier. The masks force the network to attend to clinically relevant zones and exclude the rest of the face, which is what yields the privacy-preserving property. The comparison of single-region vs multi-region configurations is what reveals complementarity among regions.

What would settle it

Collect a larger, clinically confirmed rosacea test set (e.g., 200+ patients) with automated face alignment, and compare full-face AUC against the best patch combination (e.g., forehead+nose+left cheek). If full-face AUC equals or exceeds the patch AUC on this set, the paper's central claim that patch inputs match or beat full-image classification is refuted for clinically representative conditions.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a ResNet-18 trained on masked facial regions detects rosacea at least as reliably as the same network trained on full-face images, and that region choice matters more than patch size. In experiments on 50 real rosacea-positive and 150 real negative face images, combinations of multiple patches (e.g., forehead + nose + left cheek) achieved area-under-the-ROC-curve values of 0.99–1.00, while single patches such as forehead or nose alone fell to about 0.88. The full-face baseline was sometimes outperformed by the best patch combination, suggesting that irrelevant facial context hurts rather than helps. Cheek patches carried the most discriminative

Load-bearing premise

The ranking of patch strategies rests on a small real test set (50 positive, 150 negative) whose labels come from web collection and whose patch placement relies on manual eye-alignment that may not transfer from synthetic to real faces.

Editorial extensions

If this is right

  • Patch-based models using cheeks plus forehead/nose can match or beat full-face models, so diagnostically relevant information is concentrated in localized facial zones.
  • Cheek regions are the strongest single source of rosacea signal, so future systems should weight them more heavily.
  • Because smaller patches retain most of the diagnostic accuracy, privacy-preserving inputs can be used without a large performance penalty.
  • Full-face inputs carry noise that can degrade classification, so cropping out non-clinical regions is a beneficial preprocessing step.
  • The approach provides a template for other region-specific dermatological screening tasks, such as acne or perioral dermatitis.

Reading between the lines

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

  • If the reported AUCs generalize, an immediate testable extension is whether automatic face alignment (rather than manual eye clicks) preserves the patch-strategy ranking on larger, multi-source clinical datasets.
  • The privacy claim is about excluding identifiable features, but skin texture and geometry in patches may still be quasi-identifying; a re-identification study on the patches alone would tell whether the privacy guarantee holds.
  • The apparent superiority of patches over full-face inputs suggests that masking acts as a strong regularizer; one could test this by adding controlled background noise to full-face inputs and observing whether AUC drops toward patch levels.
  • The synthetic-to-real gap in validation vs test AUC implies the training distribution is optimistic; an external validation with real patient photos is the natural next step before any clinical use.
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

4 major / 5 minor

Summary. The paper proposes patch-based rosacea detection using a ResNet-18 fine-tuned on synthetic face images and evaluated on a real-image test set. Four facial ROIs (forehead, nose, left cheek, right cheek) are combined singly and in multi-region configurations at three manual scales (small, medium, large), and results are compared against full-face input. The main empirical claim, stated in the abstract and conclusion, is that several patch-based strategies achieve competitive or superior accuracy and sensitivity relative to full-image baselines, while also improving privacy by excluding identifiable facial features. Results are presented as ROC curves and AUC values on a synthetic validation set and a 200-image real test set (50 rosacea-positive, 150 negative).

Significance. If substantiated, the work would provide a practical, privacy-conscious alternative to full-face classification for rosacea screening, and the systematic comparison of patch locations and sizes would be useful guidance for dermatological AI. The paper's strengths are the breadth of evaluated patch configurations, the inclusion of a real-world test set, and a clear motivation for region-focused diagnosis. However, the central claim currently rests on under-powered, unreplicated AUC comparisons with no confidence intervals, no significance testing, and a small, label-unverified positive test set; the reported evidence is insufficient to establish superiority over full-face input.

major comments (4)
  1. [Abstract, §4.3, §5] The central claim that patch strategies achieve 'competitive or superior accuracy and sensitivity' is not supported by the reported results. Section 4.3 reports only ROC curves and AUC values; no numeric accuracy or sensitivity values are given anywhere, despite the abstract explicitly promising them. With only 50 positive test images, the standard error of an AUC near 0.95 is about 0.03, so differences such as 'F+N+LC at 1.00 AUC' versus a lower full-face AUC can easily arise by chance. The authors should report sensitivity/specificity at a clinically relevant threshold, bootstrap confidence intervals for AUC, and a paired significance test (e.g., DeLong) for the key comparisons, including the full-face baseline values that are currently not tabulated.
  2. [§4.1, §4.3] The reliability of the test set is not established. The 50 rosacea-positive images are web-collected from Kaggle, DermNet, and the National Rosacea Society without clinician verification, and the 150 CelebA 'rosacea-negative' images may contain undiagnosed cases. If label noise is differential across the patch configurations, the ranking of AUCs could be an artifact. The paper should describe inclusion/exclusion criteria, preferably include independent dermatologist review of the positive labels and a screening of negatives, and state whether any images were excluded and why.
  3. [§4.1, §4.3] The evaluation ignores multiple comparisons and the synthetic validation distribution. Roughly 15 configurations times 3 scales yields about 45 AUC values, and the best-performing combinations are highlighted without any correction or pre-registered hypothesis. Moreover, the validation set is drawn from the same GAN-generated distribution as the training set (the paper itself notes 'perfect validation performance is largely due to the high similarity'), so model selection by validation accuracy does not provide evidence of generalization. The authors should report all configuration results in a table, apply a multiple-testing correction or clearly label exploratory findings, and evaluate on additional unseen synthetic or real data.
  4. [§3.1, §3.2, §4.2] The patch extraction procedure is not reproducible as described. The three patch scales ('small, medium, large') and the ROI masks are defined only qualitatively, and the input to the network is described as a 150×130×3 aligned crop with a mask applied; it is unclear whether masked regions are zeroed, mean-filled, or otherwise handled, and whether patch 'size' changes the spatial extent of the input or merely the mask. The eye coordinates and ROI boundaries are selected manually, so the paper should provide concrete pixel coordinates or an automatic alignment method, and quantify the sensitivity of results to alignment variability.
minor comments (5)
  1. [Throughout] There are numerous typos and grammatical errors, e.g., 'requirs', 'pateches', 'learing', 'The networks was', 'Epochs' in Table 1. A thorough proofread is needed.
  2. [References] References [3] and [15] are the same Campanella et al. article and should be merged or distinguished.
  3. [§3.2] The viewer-frame definition of LC/RC is useful, but the parenthetical 'subject's anatomical right' for image-left cheek is confusing and should be illustrated or clarified.
  4. [§4.2] The paper reports a single training configuration (SGD, LR 1e-3, step decay). It would be helpful to state whether any hyperparameter tuning was performed and how the final configuration was selected, since all configurations share the same hyperparameters.
  5. [Figures] The paper references Figures 1–4 but they are not included in the text provided. In the final version, ROC curves should be legible and each configuration labeled; a table of AUC values with confidence intervals would be more informative than curves alone.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the patch-superiority claim is an empirical test-set result, not equivalent to the paper's inputs or self-citations.

full rationale

The paper makes no first-principles derivation that could reduce to its inputs. Its central claim, that several patch-based strategies achieve competitive or superior accuracy and sensitivity than full-image methods, is supported by held-out real-world test AUCs in Section 4.3 (50 rosacea-positive images from Kaggle/DermNet/National Rosacea Society and 150 negatives from CelebA). These test images are external to training, and the model is not fitted to them. The self-citations [7][8] are used only as data provenance: Section 4.1 states 'we use the same datasets as [7][8] that are synthetically generated from GANs.' That is a pointer to how the training/validation data were created, not a load-bearing argument for the patch-strategy conclusion. The paper explicitly acknowledges that its near-perfect validation performance is 'largely due to the high similarity between the training and validation sets' and therefore evaluates generalization on real data, which is the correct non-circular practice. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported, and no ansatz is smuggled in via citation. The absence of confidence intervals or significance testing is a statistical robustness concern, not a circularity concern.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a chain of domain assumptions: synthetic training data transfers to real clinical images, the four ROIs capture all relevant cues, CelebA is a valid negative population, and cropping removes identity. None of these is validated with external data or privacy tests. The only tunable choices disclosed are manual patch scales, eye alignment, and standard training hyperparameters.

free parameters (3)
  • ROI patch scales (small, medium, large) = not disclosed (manual)
    Three scales are generated manually (Section 3.1) with no quantitative mask definitions; all performance results depend on these unstated choices.
  • eye alignment coordinates = not disclosed
    Faces are aligned by manually selecting eye coordinates (Section 4.1); alignment errors would affect every patch configuration and the synthetic-to-real transfer.
  • training hyperparameters (learning rate, momentum, step size, epochs) = eta=1e-3, momentum=0.9, step_size=7, epochs=30
    These are standard fine-tuning choices (Table 1), not fitted to the central comparison, but they are single-run choices and no seed or repeated-run analysis is reported.
assumptions (4)
  • domain assumption Synthetic GAN-generated rosacea images are representative of real rosacea facial patterns
    Training and validation use GAN-generated positives (Section 4.1, refs [9][10]). If the synthetic distribution differs from real disease appearance, the learned features may not transfer to the real test set.
  • domain assumption The four ROIs (forehead, nose, left cheek, right cheek) contain all clinically relevant rosacea cues
    Patch masks limit the input to these regions (Section 3.1). Any diagnostic signal outside them is discarded by construction, so the central performance comparison assumes these are the right regions.
  • domain assumption CelebA images are valid rosacea-negative controls
    The 150 negative test images come from CelebA (Section 4.1), a general face dataset not screened for rosacea, so label noise is possible.
  • domain assumption Cropping to patches removes personally identifiable information
    The privacy claim rests on the assumption that localized patches cannot identify a person; no re-identification or privacy evaluation is provided in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Patch-based Automatic Rosacea Detection Using the ResNet Deep Learning Framework." pith.science (2026). https://pith.science/paper/5FACXP62

@misc{pith2026250909841,
  author       = {Pith},
  title        = {Pith review of: Patch-based Automatic Rosacea Detection Using the ResNet Deep Learning Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5FACXP62}},
  note         = {Machine review of arXiv:2509.09841}
}
read the original abstract

Rosacea, which is a chronic inflammatory skin condition that manifests with facial redness, papules, and visible blood vessels, often requirs precise and early detection for significantly improving treatment effectiveness. This paper presents new patch-based automatic rosacea detection strategies using the ResNet-18 deep learning framework. The contributions of the proposed strategies come from the following aspects. First, various image pateches are extracted from the facial images of people in different sizes, shapes, and locations. Second, a number of investigation studies are carried out to evaluate how the localized visual information influences the deep learing model performance. Third, thorough experiments are implemented to reveal that several patch-based automatic rosacea detection strategies achieve competitive or superior accuracy and sensitivity than the full-image based methods. And finally, the proposed patch-based strategies, which use only localized patches, inherently preserve patient privacy by excluding any identifiable facial features from the data. The experimental results indicate that the proposed patch-based strategies guide the deep learning model to focus on clinically relevant regions, enhance robustness and interpretability, and protect patient privacy. As a result, the proposed strategies offer practical insights for improving automated dermatological diagnostics.

Figures

Figures reproduced from arXiv: 2509.09841 by the authors.

Figure 1
Figure 1. An example of patch selection with different sizes. From left to right: mean face, large size patches, medium size patches, small size patches [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An example of different patch combinations for a face with rosacea. First row from left to right: full face, forehead (F) alone, the nose (N) alone, the left cheek (LC) alone, the right cheek (RC) alone, F+N, F+LC, F+RC. Second row from left to right: N+LC, N+RC, F+N+LC, F+N+RC, F+LC+RC, N+LC+RC and all four patches. 3.3 Classification Model We used ResNet-18, a residual convolutional neural network with 18 layers t… view at source ↗
Figure 3
Figure 3. The first column shows the mean images of the two training datasets from the rosacea negative and positive classes, respectively. The remaining seven columns display seven pair of random example images from the two training datasets corresponding to the normal people and those with rosacea, respectively. Training and Validation For rosacea positive cases, 300 frontal face images synthesized using a GAN framework [9]… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The ROC curves on validation and test set for different patch size configurations [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 3 canonical work pages

  1. [1]

    Clinical, Cosmetic and Investigational Dermatology, 1465–1473 (2022)

    Ge, L., Li, Y., Wu, Y., Fan, Z., Song, Z.: Differential diagnosis of rosacea using ma- chine learning and dermoscopy. Clinical, Cosmetic and Investigational Dermatology, 1465–1473 (2022)

  2. [2]

    American Journal of Clinical Dermatology22(4), 457–465 (2021).https://doi

    van Zuuren, E.J., Arents, B.W.M., van der Linden, M.M.D., Vermeulen, S., Fe- dorowicz, Z., Tan, J.: Rosacea: New Concepts in Classification and Treatment. American Journal of Clinical Dermatology22(4), 457–465 (2021).https://doi. org/10.1007/s40257-021-00595-7

  3. [4]

    Entropy21(2), 168 (2019).https://doi

    Wang, C., Zhao, Z., Ren, Q., Xu, Y., Yu, Y.: Dense U-Net Based on Patch-Based Learning for Retinal Vessel Segmentation. Entropy21(2), 168 (2019).https://doi. org/10.3390/e21020168

  4. [5]

    In: Proc

    He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pp. 770–778 (2016).https://arxiv.org/abs/1512.03385

  5. [6]

    International Journal of Computer Vision115, 211–252 (2015).https: //doi.org/10.1007/s11263-015-0816-y

    Russakovsky, O., Deng, J., Su, H., et al.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision115, 211–252 (2015).https: //doi.org/10.1007/s11263-015-0816-y

  6. [7]

    In: Proc

    Yang, C., Liu, C.: Increasing rosacea awareness among population using deep learning and statistical approaches. In: Proc. Int. Conf. on Medical Imaging and Computer-Aided Diagnosis, pp. 110–119. Springer (2024)

  7. [8]

    arXiv preprint arXiv:2504.08073 (2025)

    Yang, C., Liu, C.: Interpretable automatic rosacea detection with whitened cosine similarity. arXiv preprint arXiv:2504.08073 (2025)

  8. [9]

    In: Proc

    Mohanty, A., Sutherland, A., Bezbradica, M., Javidnia, H.: Towards synthetic gen- eration of clinical rosacea images with GAN models. In: Proc. Irish Signals and Systems Conf., pp. 1–5 (2022) 10 Chengyu et al

Show all 17 references
  1. [10]

    Communications of the ACM63(11), 139–144 (2020)

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM63(11), 139–144 (2020)

  2. [11]

    Nature 542(7639), 115–118 (2017)

    Esteva, A., Kuprel, B., Novoa, R.A., Ko, J., Swetter, S.M., Blau, H.M., Thrun, S.: Dermatologist-level classification of skin cancer with deep neural networks. Nature 542(7639), 115–118 (2017)

  3. [12]

    Scientific Data5(1), 1–9 (2018)

    Tschandl, P., Rosendahl, C., Kittler, H.: The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data5(1), 1–9 (2018)

  4. [13]

    arXiv preprint arXiv:2411.09863 (2024).https://arxiv.org/abs/2411.09863

    Cao, J., Chen, X., Liu, B., Ding, M., Xie, R., Song, L., Li, Z., Zhang, W.: Face De- identification: State-of-the-art Methods and Comparative Studies. arXiv preprint arXiv:2411.09863 (2024).https://arxiv.org/abs/2411.09863

  5. [14]

    Scientific Reports11, 13524 (2021).https://doi.org/10.1038/ s41598-021-93030-0

    Ziller, A., Usynin, D., Braren, R., et al.: Medical imaging deep learning with dif- ferential privacy. Scientific Reports11, 13524 (2021).https://doi.org/10.1038/ s41598-021-93030-0

  6. [15]

    Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical- grade computational pathology using weakly supervised deep learning on whole slideimages.NatureMedicine25(8),1301–130...

  7. [16]

    Huynh, H.N., Tran, A.T., Tran, T.N.: Region-of-Interest Optimization for Deep- Learning-BasedBreastCancerDetectioninMammograms.AppliedSciences13(12), 6894 (2023).https://doi.org/10.3390/app13126894

  8. [17]

    In: Proc

    Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pp. 4401–4410 (2019)

  9. [18]

    In: Proc

    Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proc. IEEE Int. Conf. on Computer Vision (ICCV), pp. 3730–3738 (2015)

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.