Pith. sign in

REVIEW 3 major objections 6 minor 26 references

White Light Specular Reflection Data Augmentation for Deep Learning Polyp Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Pasting synthetic white-light reflections into colonoscopy training images helps polyp detectors stop confusing glare with polyps, raising a standard detection score from 77.5% to 80.7%.

desk verdict A simple, new reflection-inpainting augmentation that probably helps a bit, but the headline gain is not statistically pinned down because the rate was chosen on the test set. read the letter →

arxiv 2505.05248 v1 pith:W6PPFOBQ submitted 2025-05-08 eess.IV cs.CV

Jose Angel Nuñez , Fabian Vazquez , Diego Adame , Xiaoyan Fu , Pengfei Gu , Bin Fu This is my paper · ORCID
classification eess.IVcs.CV
keywords polypdetectiondataaugmentationspecularreflectionwhitelightcolonoscopyYOLOv5deeplearningfalsepositives
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

The paper is trying to establish that adding artificial white-light specular reflections to colonoscopy training images makes deep-learning polyp detectors better at telling real glare from polyps. The idea is to give the model more chances to be confused during training so it learns to suppress those confusers at test time. In the main comparison, training a YOLOv5 detector with 20% of images replaced by WLSR-augmented versions raises the mean-average-precision score mAP50 from 77.5% to 80.7%, a 3.2-point gain. If the claim holds, data augmentation alone can reduce a specific class of false positives without new network architectures or manual relabeling.

What carries the argument

The load-bearing mechanism is a three-stage data-augmentation pipeline. First, a Bank of Lights is created by cropping 300 real white-light reflection regions from the training set, adjusting their intensity shades, and applying flips, random scaling between 0.8 and 1.2, and random rotation between -30 and 30 degrees, yielding 3,600 patches. Second, prohibited areas are identified and colored orange: existing lights, polyp bounding boxes, and black borders in the outer 20% margins. Third, a sliding window moves across each image in steps of the patch width plus one and height plus one, records every top-left position whose covered area contains no orange pixels, and randomly places one artificial light at one such position. This ensures the added glare looks natural, does not occlude polyps, and does not deform existing reflections.

What would settle it

Run the same 20%-replacement comparison on a colonoscopy test set recorded with a different endoscope model or light source, keeping the synthetic reflection bank fixed; if the mAP50 gain disappears while the bank still looks similar to training-set reflections, the augmentation is not transferring to real reflections.

Watch

Extended reading notes

Core claim

The central claim is that deliberately increasing the number of white-light specular reflections in training images improves polyp detection, rather than hurting it. The paper builds a bank of 3,600 artificial light patches from real reflection regions, marks areas where a new light must not go (existing lights, polyps, and black borders), and pastes one synthetic light into each eligible training image via a sliding-window search. Training YOLOv5 with 20% WLSR-replaced images yields 80.7% mAP50 versus 77.5% with the original images, and recall rises from 63.6% to 71.5%. Combining WLSR with standard augmentations gives 81.5% mAP50, suggesting the augmentation is additive. The authors interpret the results as confirming that harder training scenarios, with more reflection opportunities to make mistakes, give the detector more chances to learn to distinguish polyps from specular highlights.

Load-bearing premise

The load-bearing premise is that synthetic reflections cropped from the training images and pasted into non-prohibited areas are realistic enough to stand in for the real white-light reflections the detector will face at test time.

Editorial extensions

If this is right

  • With a 20% replacement rate, WLSR raises mAP50 by 3.2 points over the original-image baseline, and recall improves by 7.9 points, so the augmentation changes detection behavior and not just confidence calibration.
  • WLSR stacks with ordinary augmentations: adding it to saturation, value, rotation, and flip augmentations raises mAP50 to 81.5%, so it can be plugged into existing training recipes.
  • Even 100% WLSR replacement still improves mAP50 slightly (78.4% versus 77.5%), which suggests the effect is not an artifact of mixing augmented and original images.
  • The visual example shows a model trained with WLSR no longer flags a specular reflection as a polyp, directly addressing one documented source of false positives in polyp detection.

Reading between the lines

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

  • Editorial inference: because the paper only pastes one light per image and finds 20% replacement optimal, testing multiple lights per image and a finer sweep of replacement rates would reveal whether the gain saturates or has a dose-response ceiling.
  • Editorial inference: the same prohibited-area and sliding-window recipe could transfer to other endoscopic and laparoscopic domains with specular highlights, but the paper reports no evidence beyond the single colonoscopy dataset it uses.
  • Editorial inference: if the benefit comes mainly from adding bright, blob-shaped distractors, then simpler bright-ellipse or brightness-spike augmentations might capture part of the effect; comparing WLSR against such generic distractors would isolate what the realism of the reflection bank contributes.
  • Editorial inference: the method does not require a segmentation network, only bounding-box annotations and a white-mask threshold, so a natural test is whether the augmentation becomes even stronger when the reflection bank is built from a learned reflection detector rather than a fixed white threshold.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a data augmentation method, White Light Specular Reflection (WLSR), which adds synthetic specular reflections to colonoscopy training images to reduce false-positive polyp detections. The method first builds a bank of light patches cropped from real training images, applies geometric and intensity transformations, identifies prohibited regions (existing lights, polyps, black borders), and pastes one synthetic light into a random non-prohibited location using a sliding-window search. Experiments on the Harvard Dataverse polyp dataset with YOLOv5 report that training with 20% WLSR replacement improves mAP50 from 77.5% to 80.7%, and combining WLSR with standard augmentations yields 81.5% mAP50. The paper concludes that WLSR is an effective plug-in augmentation.

Significance. If the reported improvement is genuine, WLSR is a simple, domain-inspired augmentation that could help polyp detectors avoid a common failure mode, and it may be applicable to other endoscopic tasks with specular reflections. The method is intuitive and described in sufficient algorithmic detail to be reimplemented. However, the current statistical evidence is weak: the headline 3.2-point gain is the best of four tested replacement rates selected on the same test set, and no measures of variance are reported. The paper would be strengthened by a held-out tuning protocol, per-seed results with standard deviations, and possibly a significance test. The core idea has merit, but the evidence presented does not yet establish the claimed improvement.

major comments (3)
  1. [Section 3, Table 1] The 20% replacement rate was selected as the best among 10%, 20%, 50%, and 100% after evaluating all configurations on the same test split. The headline 3.2% mAP50 gain is therefore the maximum of four comparisons rather than a pre-specified estimate, which introduces optimistic selection bias. Without a held-out tuning set or a correction for multiple comparisons, the central claim that WLSR at 20% replacement yields 80.7% mAP50 is not adequately supported. The authors should either pre-register the rate, tune it on a validation split and report test performance only for the chosen rate, or provide confidence intervals that account for the selection.
  2. [Section 3, Experimental Results and Table 1] The paper reports only averaged metrics over three seeds (seeds 1, 2, and 3) and gives no standard deviations, per-seed values, confidence intervals, or significance tests. Given that the other replacement rates (10%, 50%, 100%) show much smaller gains (0.7, 1.6, and 0.9 percentage points) and deep learning training is typically noisy, the 3.2-point gap could plausibly lie within run-to-run variability. The authors should report the full distribution of results and ideally a paired significance test (e.g., a paired t-test or Wilcoxon signed-rank test over seeds) to support the claim.
  3. [Section 2.1] The generation of the light bank relies on a binary mask with 'a threshold range for shades of white' (Section 2.2), but the actual threshold values are not specified anywhere in the manuscript. This threshold is a free parameter that directly controls the appearance of the synthetic reflections and thus the difficulty of the augmentation. The authors should state the threshold range and, preferably, provide a visual or quantitative comparison between the distribution of generated lights and real reflections to support the claim of realism.
minor comments (6)
  1. [Section 3, Experimental Results with 100% Replacement] '0.90% improved in mAP50' should read '0.9 percentage points improvement in mAP50' (similarly '3.2% improvement' is a percentage-point change).
  2. [Section 2.2] There is a typo: 'procedures;in this case' should be 'procedures; in this case'.
  3. [Section 2.3] The retry count n mentioned in 'the algorithm tries again n number of times' is never given a numeric value; please specify the value used in the experiments.
  4. [Section 3, Implementation Details] There is a typo: 'equiped' should be 'equipped'.
  5. [Section 1] In the introduction, 'two-stage detectors. such as Faster R-CNN' contains an erroneous period after 'detectors'.
  6. [Section 2.3] The sliding-window step size is described as the width/height of the cropped light image plus one, but it is not clear whether the overlap is intended or whether the step size should be exactly the width/height to avoid overlapping windows; clarify the design choice.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: the WLSR claim is an empirical augmentation result tested on a held-out split, with no derivation or fitted quantity that collapses into its own input.

full rationale

This paper contains no mathematical derivation chain to audit; the central claim is an empirical one about data augmentation. The artificial light bank is generated from the training images (Section 2.1), but the augmented training set is then used to train a detector that is evaluated on the separate test 2019 split, so the reported gains (Table 1) are not equal to the construction inputs by definition. No parameter is fitted to the test set and then renamed a prediction; the replacement percentages are experimental configurations, and the 20% choice is reported as the best among those tested rather than a fitted quantity. The paper includes self-citations (e.g., refs. [3]-[5]) but they support general medical-imaging context and are not load-bearing for the WLSR claim. The concern that the 20% replacement rate was selected after observing test results, and the absence of variance or significance reporting, is a selection-bias/statistical-reliability issue, not circularity under the specified definitions. Consequently, no circular step can be quoted with a specific reduction, and the honest finding is a score of 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

There is no mathematical derivation to audit; the empirical claim rests on manual design choices (replacement rate, retry count, threshold, transform ranges) and on the domain assumption that synthetic reflections transfer to real colonoscopy images. No code or data artifacts are shipped, so the reader must supply or guess several parameters to reproduce the result.

free parameters (5)
  • Replacement percentage = 20%
    The best of 10%, 20%, 50%, and 100% replacement rates, selected by test-set performance in Table 1. This is a hyperparameter fitted to the evaluation data.
  • Number of synthetic lights per augmented image = 1
    The authors chose one light to see if any improvement appears; this manual choice limits the method's search space and is not justified by experiments.
  • White-mask threshold range = unspecified
    Section 2.2 identifies light regions using a threshold range for shades of white, but the actual values are not given. This parameter controls which real reflections are cropped into the bank.
  • Retry count n for light placement = unspecified
    Section 2.3 says an image is excluded after n tries, but n is never defined. This affects which images are kept for training and could influence final performance.
  • Transform ranges = scale 0.8-1.2; rotation -30 to 30 degrees
    Manual choices in Section 2.1 determine the diversity of the synthetic light bank; no sensitivity analysis is provided.
assumptions (4)
  • domain assumption Adding synthetic specular reflections to training images will make the detector robust to real reflections at test time.
    Central to the method's mechanism; the mAP gain in Section 3 is interpreted as evidence that the model learned to distinguish polyps from reflections, but this assumes the synthetic lights match the test-time artifact distribution.
  • domain assumption Orange is absent from real colonoscopy images.
    Section 2.2 uses orange to mark prohibited areas. If orange appears in real images, the sliding-window criterion could behave differently at test time.
  • domain assumption Pre-trained COCO weights provide a useful starting point for polyp detection.
    All models initialize from a COCO pre-trained model; the paper treats this as standard practice without testing its influence on the augmentation effect.
  • domain assumption Default YOLOv5 hyperparameters are adequate and allow fair comparison.
    The authors use default training parameters for all experiments, assuming these are suitable for polyp detection and do not interact negatively with WLSR.

how reviews work

0 comments
Cite this review

Pith. "Pith review of White Light Specular Reflection Data Augmentation for Deep Learning Polyp Detection." pith.science (2026). https://pith.science/paper/W6PPFOBQ

@misc{pith2026250505248,
  author       = {Pith},
  title        = {Pith review of: White Light Specular Reflection Data Augmentation for Deep Learning Polyp Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W6PPFOBQ}},
  note         = {Machine review of arXiv:2505.05248}
}
read the original abstract

Colorectal cancer is one of the deadliest cancers today, but it can be prevented through early detection of malignant polyps in the colon, primarily via colonoscopies. While this method has saved many lives, human error remains a significant challenge, as missing a polyp could have fatal consequences for the patient. Deep learning (DL) polyp detectors offer a promising solution. However, existing DL polyp detectors often mistake white light reflections from the endoscope for polyps, which can lead to false positives.To address this challenge, in this paper, we propose a novel data augmentation approach that artificially adds more white light reflections to create harder training scenarios. Specifically, we first generate a bank of artificial lights using the training dataset. Then we find the regions of the training images that we should not add these artificial lights on. Finally, we propose a sliding window method to add the artificial light to the areas that fit of the training images, resulting in augmented images. By providing the model with more opportunities to make mistakes, we hypothesize that it will also have more chances to learn from those mistakes, ultimately improving its performance in polyp detection. Experimental results demonstrate the effectiveness of our new data augmentation method.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 24 canonical work pages

  1. [1]

    White Light Specular Reflection Data Augmentation for Deep Learning Polyp Detection

    INTRODUCTION Colorectal cancer (CRC) is a significant global health issue, which is the third most common cancer worldwide in 2021 and the second leading cause of cancer-related deaths [1, 2, 3, 4, 5]. The key to preventing CRC is the early detection and removal of adenomatous polyps. Colonoscopies are known as the gold standard for detecting polyps [6, 7...

    work page Pith review arXiv 2021

  2. [2]

    2 shows the pipeline of our new data augmentation method: First, we generate a bank of artificial lights using the training dataset (Section 2.1)

    METHODOLOGY Fig. 2 shows the pipeline of our new data augmentation method: First, we generate a bank of artificial lights using the training dataset (Section 2.1). Then for input training im- ages, we identify the regions in which we cannot place these lights (Section 2.2). Next, we place these artificial lights to the areas that fit to form the augmented...

    work page

  3. [3]

    Each of the 4 transformations listed above transforms the original 300 re- gions: thus, 4 x 300 = 1,200 new images

    random scaling (scaling between 0.8 to 1.2), and 4) ran- dom rotation (rotation range: -30 to 30 degrees). Each of the 4 transformations listed above transforms the original 300 re- gions: thus, 4 x 300 = 1,200 new images. In the second round, the images flipped vertically and horizontally get transformed by random scaling and random rotation to create mo...

    work page

  4. [4]

    The data used for this research comes from the Harvard Dataverse [19]

    EXPERIMENTS AND RESULTS Dataset. The data used for this research comes from the Harvard Dataverse [19]. This dataset comprises polyp im- ages sourced from MICCAI 2017, CVC Colon DB, GLRC, KUMC, and the University of Kansas Medical Center, each accompanied by XML annotations. The XML annotations describe important details such as coordinates of where the p...

    work page 2017

  5. [5]

    CONCLUSION In this paper, we presented a new data augmentation named WLSR for DL polyp detection. Our method is capable of adding more specular reflections to training images to make training harder and hence to give the model more chances to learn new features that would help it distinguish a polyp from a specular reflection. Experimental results on one ...

    work page

  6. [6]

    Ethical approval was not required as confirmed by the licenses attached with the open access datasets

    COMPLIANCE WITH ETHICAL STANDARDS This research study was conducted retrospectively using hu- man subject data made available in open access by one pub- licly available dataset [19]. Ethical approval was not required as confirmed by the licenses attached with the open access datasets

    work page

  7. [7]

    ACKNOWLEDGEMENTS This work was supported in part by the Graduate Assistance in Areas of National Need (GAANN) grant

    work page

  8. [8]

    Colorectal cancer statistics, 2023,

    R. L. Siegel, N. S. Wagle, A. Cercek, R. A. Smith, and A. Jemal, “Colorectal cancer statistics, 2023,” CA: A Cancer Journal for Clinicians , vol. 73, no. 3, pp. 233– 254, 2023

    work page 2023

Show all 26 references
  1. [9]

    Colorectal cancer: Key facts,

    W. H. Organization, “Colorectal cancer: Key facts,” https://www.who.int/news-room/ fact-sheets/detail/cancer, 2021, Accessed: 22-Oct-2024

    2021

  2. [10]

    IHCSurv: Ef- fective immunohistochemistry priors for cancer survival analysis in gigapixel multi-stain whole slide images,

    Y . Zhang, H. Chao, Z. Qiu, W. Liu, Y . Shen, N. Sapkota, P. Gu, D. Z. Chen, L. Lu, K. Yan, et al., “IHCSurv: Ef- fective immunohistochemistry priors for cancer survival analysis in gigapixel multi-stain whole slide images,” in MICCAI, 2024, pp. 211–221

    2024

  3. [11]

    Boosting medical im- age classification with segmentation foundation model,

    P. Gu, Z. Zhao, H. Wang, Y . Peng, Y . Zhang, N. Sap- kota, C. Wang, and D. Z. Chen, “Boosting medical im- age classification with segmentation foundation model,” arXiv preprint arXiv:2406.11026, 2024

    2024 arXiv

  4. [12]

    Path-GPTOmic: A balanced multi-modal learn- ing framework for survival outcome prediction,

    H. Wang, Y . Yang, Z. Zhao, P. Gu, N. Sapkota, and D. Z. Chen, “Path-GPTOmic: A balanced multi-modal learn- ing framework for survival outcome prediction,” arXiv preprint arXiv:2403.11375, 2024

    2024 arXiv

  5. [13]

    The impact of screening on colorectal cancer mortality and incidence: has it really made a dif- ference?,

    A. G. Zauber, “The impact of screening on colorectal cancer mortality and incidence: has it really made a dif- ference?,” Digestive Diseases and Sciences, vol. 60, pp. 681–691, 2015

    2015

  6. [14]

    Colonic polyps: diagnosis and surveillance,

    M. B. Huck and J. L. Bohl, “Colonic polyps: diagnosis and surveillance,” Clinics in Colon and Rectal Surgery , vol. 29, no. 04, pp. 296–305, 2016

    2016

  7. [15]

    A robust real-time deep learning based automatic polyp detection system,

    I. Pacal and D. Karaboga, “A robust real-time deep learning based automatic polyp detection system,”Com- puters in Biology and Medicine , vol. 134, pp. 104519, 2021

    2021

  8. [16]

    Bernal and A

    J. Bernal and A. Histace, Computer-Aided Analysis of Gastrointestinal Videos, 2021

    2021

  9. [17]

    Colorectal polyp detection in colonoscopy images using YOLO-V8 network,

    M. Lalinia and A. Sahafi, “Colorectal polyp detection in colonoscopy images using YOLO-V8 network,” Signal, Image and Video Processing , vol. 18, no. 3, pp. 2047– 2058, 2024

    2024

  10. [18]

    Intracker: An integrated detector-tracker framework for cell detection and tracking,

    P. Liang, J. Chen, Y . Zhang, H. Wang, H. Zheng, P. Gu, and D. Chen, “Intracker: An integrated detector-tracker framework for cell detection and tracking,” in CBMS, 2020, pp. 332–337

    2020

  11. [19]

    You only look once: Unified, real-time object detection,

    J. Redmon, “You only look once: Unified, real-time object detection,” in CVPR, 2016

    2016

  12. [20]

    Faster R-CNN: Towards real-time object detection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards real-time object detection with region proposal networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 39, no. 6, pp. 1137–1149, 2016

    2016

  13. [21]

    Spatial pyramid pooling in deep convolutional networks for visual recog- nition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Spatial pyramid pooling in deep convolutional networks for visual recog- nition,” IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, vol. 37, no. 9, pp. 1904–1916, 2015

    1904

  14. [22]

    Polyp detection during colonoscopy using a regression- based convolutional neural network with a tracker,

    R. Zhang, Y . Zheng, C. C. Poon, D. Shen, and J. Y . Lau, “Polyp detection during colonoscopy using a regression- based convolutional neural network with a tracker,”Pat- tern Recognition, vol. 83, pp. 209–219, 2018

    2018

  15. [23]

    Real- time polyp detection model using convolutional neural networks,

    A. Nogueira-Rodr ´ıguez, R. Dom ´ınguez-Carbajales, F. Campos-Tato, J. Herrero, M. Puga, D. Remedios, L. Ri- vas, E. S ´anchez, A. Iglesias, J. Cubiella, et al., “Real- time polyp detection model using convolutional neural networks,” Neural Computing and Applications , vol. 34...

    2022

  16. [24]

    Yolov3: An incremental improvement,

    A. Farhadi and J. Redmon, “Yolov3: An incremental improvement,” in CVPR, 2018

    2018

  17. [25]

    Improved polyp detection from colonoscopy im- ages using finetuned YOLO-v5,

    P. Ghose, A. Ghose, D. Sadhukhan, S. Pal, and M. Mi- tra, “Improved polyp detection from colonoscopy im- ages using finetuned YOLO-v5,” Multimedia Tools and Applications, vol. 83, no. 14, pp. 42929–42954, 2024

    2024

  18. [26]

    Replication Data for: Colonoscopy Polyp Detection and Classification: Dataset Creation and Comparative Evaluations,

    G. Wang, “Replication Data for: Colonoscopy Polyp Detection and Classification: Dataset Creation and Comparative Evaluations,” 2021

    2021

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.