Pith. sign in

REVIEW 3 major objections 4 minor 22 references

The iToBoS dataset: skin region images extracted from 3D total body photographs for lesion detection

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper introduces the iToBoS dataset: 16,954 region-level skin images tiled from 3D total-body photographs of 100 participants, each annotated with lesion bounding boxes and metadata, publicly released to train and benchmark…

desk verdict A genuinely useful region-level skin lesion detection dataset, but the annotation protocol has an internal contradiction and no quantitative validation of label quality. read the letter →

arxiv 2501.18270 v1 pith:U3YE62YN submitted 2025-01-30 eess.IV cs.AIcs.CV

Anup Saha , Joseph Adeola , Nuria Ferrera , Adam Mothershaw , Gisele Rezze , Séraphin Gaborit , Brian D'Alessandro , James Hudson
show 9 more authors
Gyula Szabó Balazs Pataki Hayat Rajani Sana Nazari Hassan Hayat Clare Primiero H. Peter Soyer Josep Malvehy Rafael Garcia
This is my paper · ORCID
classification eess.IVcs.AIcs.CV
keywords skinlesiondetection3DtotalbodyphotographydatasetreleaseboundingboxannotationVECTRAWB360cancerscreeningiToBoSchallengedeeplearning
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 sets out to close a gap in skin-cancer AI data: most public datasets are dermoscopic images of single, isolated lesions centered in the frame, so models never learn what normal surrounding skin looks like. The authors present the iToBoS dataset, 16,954 high-resolution image tiles of skin regions from 100 participants, captured with the VECTRA WB360 3D total-body photography system at two clinical sites in Spain and Australia. Each tile covers roughly a 7 × 9 cm patch of skin, every suspicious lesion is marked with a bounding box, and metadata give anatomical location, age group, and a sun damage score for each image. The dataset is released publicly for the iToBoS-2024 lesion detection challenge, with the stated goal of letting researchers train and benchmark detectors on context-rich clinical images rather than isolated lesion crops, ultimately supporting early skin-cancer detection in non-clinical settings. The paper's contribution is the resource itself—its construction, privacy safeguards, and statistical characterisation—not a trained detection model.

What carries the argument

The mechanism that carries the whole contribution is the tiling-and-annotation pipeline. The VECTRA WB360 scanner's 92 cameras reconstruct a full-body 3D avatar; a dedicated tiling tool divides the avatar surface into overlapping 2D tiles of about 1012 × 827 pixels with a 45-pixel overlap, tagging each tile with an anatomical region and omitting the head to protect anonymity. On the V7 Darwin annotation platform, a trained team of medical students, nurses, and doctors draws bounding boxes around every lesion at least 2.5 mm across, assigns a sun damage score from 1 to 3, and tags tattoos and distinguishing marks for inpainting, with reviewers and then dermatologists re-checking each tile. Finally, a hierarchical stratified sampler that borrows the Wallace rule of nines allocates tiles to train and test splits, balancing anatomical region, sex-at-birth, and lesion presence while preserving rare under-30 participants. This pipeline is what converts raw total-body scans into a ground-truth detection dataset.

What would settle it

Re-annotate a random sample of about 200 tiles with two independent dermatologists who have not seen the original labels, following the same 2.5 mm minimum-size rule, and measure lesion-level agreement (matched by overlap) against the published boxes. If agreement is markedly lower for lesions under 3 mm or in high-sun-damage tiles, the ground truth would be too noisy to support reliable benchmarking; the same check on tiles that share the 45-pixel overlap could show whether duplicated lesions are labelled consistently.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that a publicly releasable, clinically useful lesion-detection dataset can be produced directly from 3D total-body photographs by tiling the reconstructed avatar into 2D skin patches. The resulting dataset has 8,473 training and 8,481 test images across 100 patients, carrying 29,403 and 30,594 bounding-box annotations respectively, built through a three-phase pipeline of capture and tiling, annotation with expert review, and hierarchically stratified public-subset selection. The authors argue that the tiles preserve the surrounding-skin context that lesion-centric datasets discard, while privacy processing—masked faces, removed 3D coordinates, randomised file names, and inpainted tattoos—makes the release ethically and legally possible. Their statistical analysis shows lesion diameters peaking near 4 mm with a right-skewed tail, bounding boxes small and consistent across splits, and a roughly 4:1 ratio of lesion-bearing to lesion-free images, which they present as evidence that the data reflect real-world clinical variability. Taken together, the paper claims this resource is suitable for training and benchmarking lesion-detection algorithms in realistic conditions.

Load-bearing premise

The load-bearing premise is that the bounding-box annotations are accurate and complete enough to serve as ground truth for training and benchmarking, yet the paper reports no inter-annotator agreement or error-rate figures and concedes in the usage notes that users may find occasional inconsistencies.

Editorial extensions

If this is right

  • Researchers can train and benchmark lesion-detection algorithms on region-level images that include surrounding skin, which the paper argues is critical context that isolated-lesion datasets lack.
  • The 8,481 test images have labels held private until the iToBoS-2024 challenge ends, so the split can serve as a common, leakage-free benchmark for comparing detectors.
  • The metadata—anatomical location, age group, and sun damage score—let models condition on demographic and environmental context, which the authors say helps distinguish lesions from healthy skin.
  • The 2.5 mm annotation threshold and sun-damage tagging mean the dataset is aimed at clinically meaningful lesion detection, including the small and subtle cases that challenge both clinicians and algorithms.
  • The stratified sampling that preserves rare cases, such as participants under 30, is intended to keep clinically significant but uncommon presentations available in both training and test sets.

Reading between the lines

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

  • An implicit test the paper leaves open: because adjacent tiles overlap by 45 pixels, some lesions appear in multiple tiles, and comparing detector outputs on those duplicates would quantify consistency and the effect of tile-boundary context.
  • The paper deliberately stops at dataset construction and statistics, so a baseline detection experiment—training a standard detector on the train split and scoring the private test labels after the challenge—is the natural next step the authors do not perform.
  • The participant pool is limited to fair skin (Fitzpatrick I–II) and predominantly European ancestry, and most annotated lesions are benign; the dataset is therefore best read as a detection benchmark for finding lesions in fair skin, not as evidence about malignancy classification or generalisation across skin types.
  • Because 3D coordinates were removed and file names randomised, the public release cannot support per-patient longitudinal tracking or body-surface-level analysis; those capabilities would require access beyond the published tiles and aggregate metadata.
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 / 4 minor

Summary. The paper describes the iToBoS dataset: 16,954 2D skin-surface tiles extracted from 3D total-body photographs of 100 participants at two clinical sites, with bounding-box annotations of skin lesions, metadata (anatomical region, age, sex-at-birth, sun damage score), and train/test splits for the iToBoS-2024 Kaggle detection challenge. The authors detail the three-phase construction pipeline (acquisition, annotation, public subset selection), report descriptive statistics on lesion size, spatial distribution, sun damage, and image dimensions, and provide access instructions, data formats, and helper code. The central claim is that the dataset provides a novel, publicly available resource for training and benchmarking lesion detection AI models.

Significance. If the bounding-box labels are reliable, the dataset fills a real gap: most public skin-image datasets are lesion-centric or crop-based, whereas iToBoS provides region-level tiles with surrounding skin context, which is important for lesion detection. The two-site collection, stratified sampling, metadata, YOLO/COCO formats, and public hosting on Kaggle with helper code are concrete strengths. The paper also documents privacy protections and ethical approvals. However, the central claim of utility as a detection benchmark is not yet fully supported: the paper provides no quantitative annotation-quality metrics (inter-annotator agreement, error-rate audit, or external validation) and no baseline detection results, and the reported lesion-size statistics appear to contradict the stated 2.5 mm minimum inclusion threshold. These issues are load-bearing because systematic label noise or protocol drift would directly bias models trained and benchmarked on the dataset.

major comments (3)
  1. [Data Analysis, Figure 9 vs. Methods (Data annotation)] The stated inclusion criterion of a 2.5 mm minimum lesion diameter is contradicted by the reported diameter statistics. Figure 9b reports minimum diameters of 0.14 mm (train) and 0.28 mm (test), and Figure 9a places 5,471 train and 4,855 test lesions in the '<3 mm' category. If the 2.5 mm threshold was enforced, lesions below 2.5 mm should be absent; if it was not enforced, the Methods description is inaccurate. In either case, the reliability of the smallest bounding boxes, which are the most difficult for detection models, is unestablished. Please clarify whether the threshold was enforced, report the number of boxes below 2.5 mm in each split, and provide an audit of the pixel-to-millimeter conversion used for these measurements.
  2. [Technical Validation] The only validation offered is the assertion that 'All annotated tiles were manually reviewed by a team of dermatologists.' For a dataset whose primary use is ground truth for lesion detection, this is insufficient. No inter-annotator agreement, no error-rate audit, no sample-level re-annotation statistics, and no external comparison (e.g., against VECTRA's automated detection or dermoscopic confirmation) are reported. Please add quantitative QA metrics, ideally broken down by lesion size, so that users can calibrate the expected label noise, particularly for small or subtle lesions.
  3. [Data Accessibility and Usage Notes] The paper claims the dataset is intended for 'training and benchmarking,' but no baseline detection results are provided and the test-set labels are private 'at least until the completion of the iToBoS lesion detection challenge.' This makes the benchmarking claim currently untestable by readers. Please include at least one reproducible baseline detector trained on the public train split, with evaluation on the test set once labels are released or on a clearly described held-out subset, and state explicitly the timeline and conditions for label release.
minor comments (4)
  1. [Figure 7] The directory structure in Figure 7 shows a 'labels' folder and labels.json for the test set, while the text states that test labels are private; please reconcile this inconsistency or annotate the figure to indicate that the test labels are placeholders or withheld.
  2. [Appendix A, Figure 11] The caption in Figure 11 begins 'Figure 11. Figure 11.'; the duplication should be removed.
  3. [Methods: 2D tile generation from 3D-TBP avatar] The product name 'VectraDBTool' appears with broken spacing ('V ectraDBTooldeveloped'); please correct the formatting.
  4. [Competing interests] The competing interests statement is a placeholder sentence ('The corresponding author is responsible for providing...'); the manuscript should include an actual statement from the authors.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: this is a dataset description with descriptive statistics and no fitted-parameter prediction chain.

full rationale

The paper is a dataset report, not a derivation or modeling claim. It describes how images were acquired, annotated, reviewed, and selected for public release, and it reports descriptive statistics of the released annotations. There are no fitted parameters that are later renamed as predictions, no quantity that is defined in terms of the quantity it is said to explain, and no load-bearing conclusion that rests solely on a self-citation. The annotation workflow is described as human labeling with clinician review, and the reported lesion-size distributions are summaries of the resulting bounding boxes, not outputs of an equation that assumes those same distributions. The only notable tension is that the stated 2.5 mm minimum lesion threshold is inconsistent with reported lesion diameters as small as 0.14 mm, but that is a potential data-quality or protocol-drift concern, not a circularity. References to prior work, including a protocol paper with overlapping authors, provide context and are not used to justify the dataset's central claim. Therefore no specific circular step can be quoted, and the appropriate finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The dataset's utility rests on several unverified domain assumptions. The most important are that visual annotations by trained annotators and dermatologists are accurate ground truth, that the 2.5 mm size threshold captures clinically relevant lesions, and that the proprietary tiling pipeline preserves geometry. The only hand-chosen numeric design parameter that directly shapes the central resource is the 2.5 mm minimum lesion diameter threshold. No invented entities are introduced.

free parameters (1)
  • minimum lesion diameter threshold = 2.5 mm
    Lesions below 2.5 mm are excluded from annotation, shaping the dataset's label distribution; this is a manual threshold based on a cited clinical study (Bono et al.), not derived from the dataset itself.
assumptions (5)
  • domain assumption Bounding boxes produced by trained annotators and reviewed by dermatologists are accurate enough to serve as ground truth for lesion detection.
    The entire utility of the dataset rests on label quality, but no inter-annotator agreement or external validation is reported (Technical Validation section).
  • domain assumption Visual inspection of 3D TBP tiles can identify 'suspicious' lesions without histopathologic confirmation, and these labels are appropriate for detection training.
    The dataset predominantly contains benign lesions by visual assessment; there is no pathology ground truth (Usage Notes section).
  • domain assumption The 2.5 mm minimum lesion diameter threshold (Bono et al. 2006) means excluded micro-lesions are clinically less significant for detection.
    This threshold shapes the label distribution and is adopted from a cited clinical study, not validated on this dataset (Data annotation section).
  • domain assumption The VECTRA WB360 and WbTilingTool produce 2D tiles with sufficient geometric fidelity and pixel spacing metadata to support lesion localization and size estimation.
    The paper relies on the proprietary toolchain's correctness without independent verification (Data Collection section).
  • domain assumption The hierarchical stratified sampling based on the Wallace rule of nines yields a representative public subset for training and testing.
    Sampling proportions are chosen by design, and outliers (age <30) are handled manually, but representativeness is not validated against an external target distribution (Public subset selection section).

how reviews work

0 comments
Cite this review

Pith. "Pith review of The iToBoS dataset: skin region images extracted from 3D total body photographs for lesion detection." pith.science (2026). https://pith.science/paper/U3YE62YN

@misc{pith2026250118270,
  author       = {Pith},
  title        = {Pith review of: The iToBoS dataset: skin region images extracted from 3D total body photographs for lesion detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U3YE62YN}},
  note         = {Machine review of arXiv:2501.18270}
}
abstract

Artificial intelligence has significantly advanced skin cancer diagnosis by enabling rapid and accurate detection of malignant lesions. In this domain, most publicly available image datasets consist of single, isolated skin lesions positioned at the center of the image. While these lesion-centric datasets have been fundamental for developing diagnostic algorithms, they lack the context of the surrounding skin, which is critical for improving lesion detection. The iToBoS dataset was created to address this challenge. It includes 16,954 images of skin regions from 100 participants, captured using 3D total body photography. Each image roughly corresponds to a $7 \times 9$ cm section of skin with all suspicious lesions annotated using bounding boxes. Additionally, the dataset provides metadata such as anatomical location, age group, and sun damage score for each image. This dataset aims to facilitate training and benchmarking of algorithms, with the goal of enabling early detection of skin cancer and deployment of this technology in non-clinical environments.

Figures

Figures reproduced from arXiv: 2501.18270 by the authors.

Figure 1
Figure 1. Three-phase methodology for dataset generation: (i) data collection through 3D-TBP imaging and 2D tile extraction, (ii) data annotation and expert review, and (iii) careful selection of a representative public subset to be released as the challenge dataset. also facilitate longitudinal tracking of lesions over time to monitor disease progression or treatment response. Canfield’s VECTRA WB360 is a state-of-the-art 3D… view at source ↗
Figure 2
Figure 2. shows an example of a 2D tiles generated from a 3D avatar. Each avatar was split into tiles with an average dimension of 1012×827 pixels (px), with a 45px overlap between adjacent tiles. This overlap was crucial for maintaining lesion visibility across tile boundaries, particularly important for lesions that might otherwise be split between tiles. Tiles at the edges and bottom of the avatar had smaller dimensions to… view at source ↗
Figure 3
Figure 3. Sample 2D tiles illustrating the six anatomic region categories: torso (anterior and posterior), left arm, right arm, left leg, right leg, and unknown regions. The 3D spatial coordinates associated with each tile were also removed to enhance privacy. This critical step ensured that the tiles could not be used to reconstruct a 3D avatar of the patient, eliminating any potential risk of re-identification through spati… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: V7 Darwin platform annotation workflow illustrating the sequential process from tile upload to clinical review. Labellers perform sun damage scoring, tattoo tagging, and lesion annotation, while reviewers verify the completeness and accuracy of all labeller tasks. Dash…
Figure 5
Figure 5. Figure 5: Examples of 2D tiles where tattoos were inpainted during the annotation phase to preserve patient privacy while maintaining skin features for lesion analysis. samples to the torso, 17% to each left and right arms, 13.5% to each left and right legs, and 0.5% to unknown …
Figure 6
Figure 6. Figure 6: Hierarchical stratified sampling workflow for balanced subset selection. The process begins by setting aside outlier cases (age < 30), then splits the remaining data by sex-at-birth, anatomical region, presence of lesions, and finally applies random sampling within eac…
Figure 7
Figure 7. Figure 7: Dataset directory structure illustrating the hierarchical organisation of image tiles, annotations (YOLO and COCO formats), and metadata files across training and test sets of the iToBoS dataset hosted on the Kaggle platform. Category Train Set Test Set Barcelona (coun…
Figure 8
Figure 8. Figure 8: Visualization of lesion annotation distributions: (a, b) Spatial distribution heatmaps of lesion centers for the train and test sets, respectively. White dashed circles (r = 0.25, centered at (0.3,0.3)) highlight a region with higher annotation density in the test set.…
Figure 9
Figure 9. Figure 9: Lesion size analysis showing both clinical categorisation and continuous size distribution characteristics. low = 1 medium = 2 high = 3 0 2,000 4,000 6,000 8,000 6,897 1,442 134 6,911 1,477 93 Sun Damage Level Number of Images Sun Damage Level Distribution - Train vs T…
Figure 10
Figure 10. Figure 10: Dataset characteristics combining clinical features (sun damage assessment) and technical specifications (image resolutions). formats (train: 25.4%, test: 35.9%). Three other dimensions (1094×890 px: train 3.9%, test 5.6%; 1090×894 px: train 7.0%, test 10.1%; and 1090…
Figure 11
Figure 11. Figure 11: Statistical analysis of the iToBoS dataset. (a) Age distribution histogram of 100 participants, with [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 17 canonical work pages

  1. [1]

    Garbe, C. et al. Skin cancers are the most frequent cancers in fair-skinned populations, but we can prevent them. Eur. J. Cancer 204, 114074, https://doi.org/10.1016/j.ejca.2024.114074 (2024)

    arXiv 2024

  2. [2]

    Conforti, C. et al. Dermoscopy and the experienced clinicians. Int. J. Dermatol. 59, 16–22 (2020)

    work page 2020

  3. [3]

    Dinnes, J. et al. Visual inspection and dermoscopy, alone or in combination, for diagnosing keratinocyte skin cancers in adults. Cochrane Database Syst. Rev. (2018)

    work page 2018

  4. [4]

    & Xie, F

    Zhang, Y ., Chen, J., Wang, K. & Xie, F. Ecl: Class-enhancement contrastive learning for long-tailed skin lesion classification. In International Conference on Medical Image Computing and Computer-Assisted Intervention , 244–254 (Springer, 2023)

    work page 2023

  5. [5]

    & Lee, W

    Kim, C., Jang, M., Han, Y ., Hong, Y . & Lee, W. Skin lesion classification using hybrid convolutional neural network with edge, color, and texture information. Appl. Sci. 13, 5497 (2023)

    work page 2023

  6. [6]

    Q., Syed, T., Marques, O

    Naqvi, M., Gilani, S. Q., Syed, T., Marques, O. & Kim, H.-C. Skin cancer detection using deep learning—a review. Diagnostics 13, 1911, https://doi.org/10.3390/diagnostics13111911 (2023)

    work page doi:10.3390/diagnostics13111911 2023

  7. [7]

    Dildar, M. et al. Skin cancer detection: a review using deep learning techniques. Int. journal environmental research public health 18, 5479 (2021)

    work page 2021

  8. [8]

    & Zeebaree, S

    Saeed, J. & Zeebaree, S. Skin lesion classification based on deep convolutional neural networks architectures. J. Appl. Sci. Technol. Trends 2, 41–51, https://doi.org/10.38094/jastt20189 (2021)

    work page doi:10.38094/jastt20189 2021

Show all 22 references
  1. [9]

    & O’Shea, S

    Ali, A.-R., Li, J. & O’Shea, S. J. Towards the automatic detection of skin lesion shape asymmetry, color variegation and diameter in dermoscopic images. Plos one 15, e0234352 (2020)

    2020

  2. [10]

    & Mandal, M

    Xu, H., Lu, C., Berendt, R., Jha, N. & Mandal, M. Automated analysis and classification of melanocytic tumor on skin whole slide images. Comput. Med. Imaging Graph. 66, 124–134, https://doi.org/10.1016/j.compmedimag.2018.01.008 (2018)

    2018 doi

  3. [11]

    & Cho, Y

    Namozov, A. & Cho, Y . I. Convolutional neural network algorithm with parameterized activation function for melanoma classification. In 2018 International Conference on Information and Communication Technology Convergence (ICTC) , 417–419, https://doi.org/10.1109/ictc.2018.853...

    2018

  4. [12]

    Codella, N. C. F. et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In2018 IEEE 15th International Symposium on Biomedical Imaging ...

    2017

  5. [13]

    & Kittler, H

    Tschandl, P., Rosendahl, C. & Kittler, H. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Sci. Data 5, https://doi.org/10.1038/sdata.2018.161 (2018)

    2018 doi

  6. [14]

    Hernández-Pérez, C. et al. Bcn20000: Dermoscopic lesions in the wild. Sci. Data 11, https://doi.org/10.1038/ s41597-024-03387-w (2024)

    2024

  7. [15]

    A patient-centric dataset of images and metadata for identifying melanomas using clinical context

    Rotemberg, V .et al. A patient-centric dataset of images and metadata for identifying melanomas using clinical context. Sci. Data 8, https://doi.org/10.1038/s41597-021-00815-z (2021)

    2021 doi

  8. [16]

    & Soyer, H

    Janda, M. & Soyer, H. P. Using advances in skin imaging technology and genomics for the early detection and prevention of melanoma. Dermatology 235, 1–3 (2018)

    2018

  9. [17]

    Rayner, J. E. et al. Clinical perspective of 3d total body photography for early detection and screening of melanoma. Front. Medicine 5, 152 (2018)

    2018

  10. [18]

    Primiero, C. A. et al. Evaluation of the efficacy of 3d total-body photography with sequential digital dermoscopy in a high-risk melanoma cohort: protocol for a randomised controlled trial. BMJ open 9, e032969 (2019)

    2019

  11. [19]

    Kurtansky, N. R. et al. The slice-3d dataset: 400,000 skin lesion image crops extracted from 3d tbp for skin cancer detection. SCIENTIFIC DATA 11, https://doi.org/10.1038/s41597-024-03743-w (2024)

    2024 doi

  12. [20]

    Primiero, C. A. et al. A protocol for annotation of total body photography for machine learning to analyze skin phenotype and lesion classification. FRONTIERS IN MEDICINE 11, https://doi.org/10.3389/fmed.2024.1380984 (2024)

    2024

  13. [21]

    Bono, A. et al. Micro-melanoma detection: a clinical study on 206 consecutive cases of pigmented skin lesions with a diameter ≤ 3 mm. Br. J. Dermatol. 155, 570–573, https://doi.org/10.1111/j.1365-2133.2006.07396.x (2006)

    2006

  14. [22]

    The exposure treatment of burns

    Wallace, A. The exposure treatment of burns. The Lancet 257, 501–504, https://doi.org/10.1016/S0140-6736(51)91975-7 (1951). 12/14 Acknowledgements This work was funded by the European Union through the iToBoS project (SC1-BHC-06-2020-965221). In addition, we extend our sincere...

    1951 doi

Pith tools

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