Pith. sign in

REVIEW 4 major objections 6 minor 31 references

Photometric augmentations that mimic real device and lighting shifts improve out-of-domain skin-lesion classification more than chasing in-domain accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 21:44 UTC pith:IVTKDPDO

load-bearing objection Solid empirical augmentation study with careful stats and unusual honesty about its own selection bias; the +0.053 OOD gain is real on the measured pool but not yet an unbiased source-level claim. the 4 major comments →

arxiv 2607.26765 v1 pith:IVTKDPDO submitted 2026-07-29 cs.CV cs.AI

Searching for Robust Augmentations to Improve Out-of-Domain Generalization in Dermoscopic Skin Cancer Classification

Alexander Kozachok , Ilya Latyshev , Evgeny Karpulevich , Elena Kozachok , Egor Ushakov , Oleg Samovarov This is my paper
classification cs.CV cs.AI
keywords dermoscopyskin cancerdomain shiftout-of-domain validationdata augmentationphotometric transformsConvNeXtROC-AUC
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Dermoscopic classifiers often fail when images come from a new clinic, dermatoscope, or lighting setup. This paper asks whether the right training-time image transforms can close that gap for a binary malignant-versus-non-malignant model. Searching single transforms, color combinations, and composite policies on multi-source public data, the authors find that photometric operations dominate the useful set, and a composite mix policy raises out-of-domain ROC-AUC by about five points on held-out HAM and ISIC sources, consistently across training seeds, while preserving in-domain performance. The practical claim is that modeling the actual sources of acquisition shift is a simple, pipeline-level lever that can matter more than architecture tweaks or in-domain score chasing—though the policy was chosen on the same held-out sources used to measure it, so the effect size still needs a source-disjoint confirmation.

Core claim

Among the policies tested, a composite mix of photometric and moderate geometric transforms delivered the largest out-of-domain gain: +0.053 ROC-AUC (95% CI +0.045 to +0.061) on an expanded pool from held-out HAM10000 and ISIC sources, with non-overlapping per-seed ranges (baseline 0.761–0.775, mix 0.806–0.829). Photometric transforms dominated the useful operations, supporting the claim that augmentations modeling real domain shift can outweigh maximizing in-domain accuracy.

What carries the argument

The mix augmentation policy: a fixed composite of flips, affine/rotate, blur/noise, geometric distortions, and especially photometric operators (ColorJitter, PlanckianJitter, HueSaturationValue, HEStain, CLAHE, and related color shifts), applied during training of a ConvNeXt-Large binary classifier and ranked by out-of-domain ROC-AUC.

Load-bearing premise

That the reported out-of-domain gain is fair evidence of generalization even though the winning policy was chosen by ranking candidates on the same held-out sources later used to measure the effect.

What would settle it

Rerun the full policy search with a source-disjoint protocol—select the policy on one archive or leave-one-source-out, then evaluate once on a completely untouched external source—and check whether mix (or any photometric-heavy policy) still beats the flip-only baseline by a comparable margin.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Training pipelines for dermoscopic triage should prioritize photometric domain-randomization before architecture changes when the deployment target is a new clinic or device.
  • In-domain ROC-AUC alone is a misleading selection criterion under source-level shift; OOD-held-out ranking changes which policy wins.
  • A simple augmentation baseline is a fair yardstick that more complex domain-adaptation or test-time methods should beat without needing target labels.
  • Clinical reading of the gain must note it is driven more by fewer false alarms on benign keratoses than by higher melanoma recall, which remains low.

Where Pith is reading between the lines

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

  • If photometric mix-style policies hold under truly source-disjoint selection, they become a one-line default for any multi-clinic dermoscopy deployment checklist.
  • The same acquisition-shift logic likely transfers to other medical imaging modalities where color cast and device preprocessing dominate domain gap (e.g., fundus photography, endoscopy).
  • Uneven per-diagnosis gains suggest pairing mix with class- or subtype-aware loss or thresholding before any screening claim is made.
  • Automated policy search optimized explicitly for source-level transfer, rather than validation accuracy, is the natural next experiment the paper points toward but does not run.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper studies whether carefully chosen data augmentations improve out-of-domain robustness of a binary malignant vs non-malignant dermoscopic classifier (ConvNeXt-Large) trained on multi-source ISIC/Derm7pt data. After screening single operators, photometric combinations, and composite policies, the authors select a mix policy that yields the largest OOD gain on held-out HAM10000 and ISIC 2019–2020. On an expanded pool from those same sources they report +0.053 ROC-AUC (95% CI +0.045 to +0.061, DeLong p<0.001), with non-overlapping ranges across four seeds (baseline 0.761–0.775, mix 0.806–0.829), and argue that photometric transforms dominate useful OOD operations. A small independent clinical set (Melanoscope) shows a single-checkpoint sensitivity rise that does not persist across seeds. The authors explicitly state that because mix was selected on the evaluation sources, the effect size is not an unbiased estimate of source-level generalization.

Significance. If the qualitative conclusion holds—that augmentations modelling acquisition-induced color/illumination shift can matter more for dermoscopic OOD transfer than maximizing in-domain accuracy—the work supplies a practical, low-friction baseline for clinical dermoscopy pipelines before heavier domain-adaptation methods. Strengths that raise the contribution above a routine ablation include lesion-ID splits, an explicit residual-overlap audit (removing overlap widens rather than shrinks the gap), image- and lesion-level bootstrap CIs, DeLong tests, multi-seed repetition with non-overlapping OOD ranges, a per-diagnosis breakdown, and unusually candid non-confirmatory language in §3.5 and §4.1. The positioning against prior in-domain augmentation rankings (Perez, Valle) and shortcut literature is clear. The main limit on significance is that the headline +0.053 is supervised target-domain model selection over eleven policies, so it cannot yet be read as an unbiased source-level effect size.

major comments (4)
  1. [§3.5, Table 7, Abstract Results] §3.5 (“What this evaluation does and does not establish”) and §4.1 correctly identify that mix was ranked on OOD ROC-AUC from the held-out HAM10000/ISIC 2019–2020 screening split, and that the expanded pool (Table 7, N=9921) is drawn from those same sources. This is supervised target-domain model selection with winner’s-curse inflation over eleven candidates. The +0.053 figure is therefore not an unbiased estimate of source-level generalization—the central quantitative claim of the abstract Results. Either (i) rerun policy selection under a source-disjoint protocol (leave-one-source-out or select on one archive and evaluate once on another), or (ii) demote +0.053 throughout abstract/results/conclusions to an explicitly provisional, selection-conditioned margin and lead with the qualitative finding (photometric ops dominate; mix is a candidate policy). As written, the abstract Results sti
  2. [§3.5–3.6, Table 7, Table 9, Table 10, Abstract] Melanoscope is the only selection-untouched external set, yet it cannot carry the generalization claim: N=472 with 22 malignant cases; single-checkpoint sensitivity 0.591→0.818 is non-significant on AUC (p=0.22) and McNemar (p=0.06), and the seed-mean difference reverses (−0.004; Table 10). Any language implying reproducible external clinical gain (including the abstract’s sensitivity sentence without immediate qualification) should be removed or tightly hedged. The paper already notes underpowering; the fix is to stop treating Melanoscope as supportive evidence of transfer and to state clearly that independent generalization remains unproven.
  3. [Table 8, §3.5, §4.1, Conclusions] Table 8 shows the OOD AUC gain is driven mainly by higher specificity on benign keratoses (e.g., pigmented benign keratosis 0.545→0.818), while recall falls for the largest malignant subgroup (“Melanoma, NOS” 0.436→0.402) and absolute melanoma recall remains 0.26–0.43 across subtypes. For a triage-oriented malignant/non-malignant framing this is load-bearing: fewer false alarms on keratoses is useful, but the paper’s clinical reading must not imply improved malignancy detection. Elevate this breakdown into the main Results narrative and the Conclusions, and qualify any triage/second-opinion framing accordingly.
  4. [§3.6, Table 10] §3.6 compares baseline and mix on disjoint seed sets ({42,1,10,20} vs {2,5,15,30}), so runs are unpaired and initialization effects are not controlled. The non-overlapping OOD ranges are still persuasive, but a shared seed set (paired ΔAUC) is a low-cost strengthening of the “consistent across seeds” claim and should be provided or the unpaired design stated as a limitation of the variance analysis.
minor comments (6)
  1. [Table 5 caption, §3.5] Table 5 screening deltas and Table 7 expanded-pool deltas are not comparable (different checkpoints and different evaluation sets). The note in §3.5 helps; add an explicit cross-reference in the Table 5 caption so readers do not treat +0.0332 and +0.053 as the same experiment.
  2. [Figure 4, §3.4] Figure 4 Grad-CAM is a single illustrative case; the text already calls it anecdotal. Consider moving it to supplementary material or adding a short multi-case panel if space allows, to avoid over-weighting one BCC example.
  3. [Table 6, §3.4] Table 6 domain-mixing indices compare ImageNet-pretrained vs mix-fine-tuned embeddings and cannot isolate augmentation from fine-tuning; the authors say so. A one-sentence pointer in the table caption would prevent mis-citation.
  4. [§2.4, Abstract Methods] Only ConvNeXt-Large is used (§2.4, §4.1). Stating in the abstract Methods that results are backbone-specific would set expectations; cross-backbone checks are appropriately left to future work but should not be implied as general.
  5. [Abstract, §2.2] Minor clarity: “predominantly source-disjoint” is well audited (§2.2), but the abstract could briefly note the 1.6% residual lesion overlap and that sensitivity analysis widened the gap, so readers do not dismiss the OOD design on leakage grounds.
  6. [Appendix B.1, Data Availability] Appendix Table A6 fully specifies mix; consider releasing the remaining policy configs (or a config repo) with the camera-ready to match the reproducibility claims in Data Availability.

Circularity Check

1 steps flagged

Material evaluation circularity: mix was selected by maximizing OOD ROC-AUC on HAM/ISIC, then the headline +0.053 is re-measured on an expanded pool from those same sources (supervised target-domain model selection / winner’s curse).

specific steps
  1. fitted input called prediction [§3.3 Table 5 (policy selection) → §3.5 Table 7 (expanded OOD evaluation); also Abstract Results/Conclusions and §4.1]
    "The mix policy was selected by ranking candidate policies on out-of-domain ROC-AUC measured on the held-out test splits of HAM10000 and ISIC 2019–2020 (Table 5), and the expanded pool used here is drawn from those same sources and contains that screening split as a subset. ... the augmentation policy is itself a top-level hyperparameter, and it was chosen using labelled data from the target out-of-domain sources. In the terminology of domain generalization this is supervised target-domain model selection, and it leaves two effects that the statistics below cannot remove: a selection bias in fa"

    mix is chosen to maximize OOD ROC-AUC on HAM/ISIC screening splits; the headline +0.053 is then the paired difference of that already-chosen policy versus baseline on an expanded pool from the identical sources. The reported ‘OOD gain’ is therefore partly the search objective itself, not an independent out-of-sample prediction. Bootstrap/DeLong tests only show the two fixed checkpoints differ on this pool—they do not undo selection bias. The paper acknowledges this and caps the claim; Melanoscope (selection-untouched) does not reproduce a seed-stable gain.

full rationale

This is an empirical augmentation-search paper, not a first-principles derivation, so there is no self-definitional equation chain, no uniqueness theorem imported from the authors, and no load-bearing self-citation. The only circularity is evaluation circularity of the fitted-input-called-prediction kind: the top-level hyperparameter (the augmentation policy) was chosen by ranking eleven candidates on out-of-domain ROC-AUC from the held-out HAM10000 and ISIC 2019–2020 splits (Table 5), and the expanded-pool gain of +0.053 (Table 7) is drawn from those same sources and contains the screening split as a subset. That makes the reported OOD effect size partly the quantity the search maximized, upward-biased by winner’s-curse over eleven policies. The paper itself states this explicitly (§3.5 ‘What this evaluation does and does not establish’; Limitations §4.1; Abstract Conclusions) and does not claim the figure is unbiased source-level generalization. Independent content remains: multi-seed non-overlap on the same pool, slight in-domain gain, residual-overlap stress tests that widen rather than shrink the gap, and single-augmentation screening that photometric ops dominate. The sole selection-untouched set (Melanoscope) does not sustain the gain across seeds. Score 5 reflects partial circularity of the central OOD effect size without definitional collapse of the whole argument.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 2 invented entities

The claim is empirical, not derived. It rests on standard ML training assumptions, a clinical binary label collapse, a predominantly source-disjoint split design, and many hand-chosen training/augmentation hyperparameters. The mix policy is a constructed object selected on target-source OOD scores. No new physical entities; the ledger is mostly domain assumptions and free knobs.

free parameters (5)
  • mix policy operator probabilities and magnitude ranges = e.g. Affine p=0.4, photometric OneOf p=0.7, ColorJitter 0.9–1.1; full table in Appendix
    Application probabilities, OneOf weights, ColorJitter/Planckian/HEStain ranges, blur limits, etc. (Table A6) are hand-designed/search-chosen and define the winning intervention.
  • policy selection criterion (OOD ROC-AUC on HAM/ISIC screening split) = mix selected (+0.0332 screening ΔAUC point estimate)
    Which of eleven composites is called best is determined by the same source family later used for the main effect size.
  • optimization and training hyperparameters = lr=8e-5, 15 epochs, batch=16
    AdamW lr 8e-5, weight decay 0.01, batch 16, 15 epochs, StepLR, focal loss, seed sets differ by policy.
  • validation-tuned decision thresholds (Youden) = baseline 0.476; mix 0.455 (main checkpoints)
    Operating-point metrics depend on thresholds fit on validation per run/policy.
  • dataset-specific RGB normalize mean/std = mean (0.658, 0.542, 0.499), std (0.237, 0.217, 0.219)
    Estimated on training data and fixed into preprocessing.
axioms (6)
  • domain assumption Binary malignant vs non-malignant label collapse (including actinic keratosis as non-malignant) is an adequate computational proxy for prioritization.
    §2.1–2.2 mapping tables; authors note it is not discharge triage and pools premalignant disease.
  • domain assumption Lesion-ID splitting plus holding out HAM10000 and ISIC 2019–2020 yields a predominantly source-disjoint OOD test despite residual MILK10k–HAM lesion_id overlap and institutional Barcelona overlap.
    §2.2 overlap audit; 158/9921 images affected; authors recompute with removal.
  • domain assumption ROC-AUC on imbalanced OOD pools is the primary comparable metric of robustness under device/protocol shift.
    Stated primary criterion §2.1; AUPRC/sensitivity reported secondarily.
  • domain assumption ImageNet-pretrained ConvNeXt-Large with classical albumentations is a sufficient backbone to rank augmentation policies.
    §2.4–2.5; single-backbone limit acknowledged in §4.1.
  • standard math Standard supervised learning, bootstrap CI, and DeLong paired ROC tests are valid for comparing retained checkpoints.
    §3.5 statistical protocol.
  • ad hoc to paper Photometric shifts in training approximate the acquisition-induced color/illumination differences that drive dermoscopic domain shift.
    Hypothesis §2.3–2.6 motivated by attribution clustering and histogram distances; HEStain included as non-native domain randomization.
invented entities (2)
  • mix composite augmentation policy no independent evidence
    purpose: Operational intervention claimed to improve OOD ROC-AUC relative to baseline.
    A specific assembled recipe (Table A5/A6), not a previously standardized benchmark policy; selected inside this search.
  • Melanoscope closed clinical external set no independent evidence
    purpose: Only evaluation set untouched by policy selection.
    Author-owned, non-public; underpowered (22 malignant) and non-persistent across seeds.

pith-pipeline@v1.2.0-daily-grok45 · 29463 in / 4154 out tokens · 113603 ms · 2026-07-30T21:44:16.184850+00:00 · methodology

0 comments
read the original abstract

Background/Objectives: Dermoscopic skin lesion classifiers often lose accuracy under domain shift across imaging devices, illumination, and capture artifacts. We study how data augmentation improves the robustness of a binary malignant-versus-non-malignant classifier, with emphasis on out-of-domain (OOD) generalization. Methods: Single augmentations, photometric combinations, and composite policies were searched on a multi-source ISIC Archive collection with Derm7pt, using a ConvNeXt-Large backbone and ROC-AUC. Splits were made at the lesion-ID level, and HAM10000 and ISIC 2019-2020 were held out as a predominantly source-disjoint OOD test. Results: The largest OOD gain came from the mix policy, and photometric transformations dominated the most useful OOD operations. On an expanded pool from the same held-out sources the gain was +0.053 (95% CI +0.045 to +0.061, p<0.001), consistent across four training seeds (per-seed ROC-AUC: baseline 0.761-0.775, mix 0.806-0.829). On a small independent clinical collection, single-checkpoint sensitivity rose from 0.591 to 0.818, but this rested on 22 malignant cases and did not persist across seeds. Conclusions: Augmentations modelling real sources of domain shift can matter more than maximizing in-domain accuracy. Because the policy was selected on the same sources used to evaluate it, a source-disjoint selection protocol is needed before this effect size can be read as unbiased.

Figures

Figures reproduced from arXiv: 2607.26765 by Alexander Kozachok, Egor Ushakov, Elena Kozachok, Evgeny Karpulevich, Ilya Latyshev, Oleg Samovarov.

Figure 1
Figure 1. Figure 1: Distribution of ISIC attribution values across the datasets. Counts are taken from the train/validation/test split files actually used in this study, so they agree with Table A1; the color scale is logarithmic and combinations that contain no images are left blank. Every dataset except the ISIC 2016–2020 group maps onto a single acquisition source, and the 179 ISIC 2020 images attributed to Hospital Clínic… view at source ↗
Figure 2
Figure 2. Figure 2: t-SNE visualization of embeddings of test images extracted from ImageNet-pretrained ConvNeXt-Large, colored by dataset (N per dataset in the legend). Partial grouping by acquisition source indicates the presence of source-specific features. t-SNE axes carry no units and are shown without ticks. Cluster analysis using K-means and hierarchical agglomerative clustering, together with distance matrices based o… view at source ↗
Figure 3
Figure 3. Figure 3: Dendrogram of the clustering of data-collection sources by color distributions and inter￾histogram distance metrics [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Grad-CAM comparison for a basal cell carcinoma case: (a) the input lesion; (b) baseline model, with attention spread over several diffuse, partly peripheral foci; (c) model trained with the selected augmentation policy, with a single compact activation centered on the lesion. Maps were computed with respect to malignancy: the classification head has a single output unit, and the saliency is the gradient of… view at source ↗
Figure 5
Figure 5. Figure 5: t-SNE projection of test-set embeddings from ConvNeXt-Large trained with the selected augmentation policy, colored by dataset, with the same palette and projection settings as [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Discrimination on the ISIC + HAM out-of-domain test (N = 9921, malignant prevalence 0.135) for the baseline and the mix policy: (a) receiver operating characteristic curves; (b) precision– recall curves, with the dashed line marking the malignant prevalence (the no-skill baseline). The mix policy improves both the area under the ROC curve (0.770 → 0.823) and the area under the precision–recall curve (0.425… view at source ↗

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

31 extracted references · 2 linked inside Pith

  1. [1]

    Global Burden of Cutaneous Melanoma in 2020 and Projections to 2040.JAMA Dermatol.2022,158, 495–503

    Arnold, M.; Singh, D.; Laversanne, M.; Vignat, J.; Vaccarella, S.; Meheus, F.; Bray, F. Global Burden of Cutaneous Melanoma in 2020 and Projections to 2040.JAMA Dermatol.2022,158, 495–503

    2020

  2. [2]

    Diagnostic Accuracy of Dermoscopy.Lancet Oncol.2002,3, 159–165

    Kittler, H.; Pehamberger, H.; Wolff, K.; Binder, M. Diagnostic Accuracy of Dermoscopy.Lancet Oncol.2002,3, 159–165

    2002

  3. [3]

    Dermatologist-Level Classification of Skin Cancer with Deep Neural Networks.Nature2017,542, 115–118

    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.Nature2017,542, 115–118

  4. [4]

    Combalia, M.; Codella, N.C.F.; Rotemberg, V .; Helba, B.; Vilaplana, V .; Reiter, O.; Carrera, C.; Barreiro, A.; Halpern, A.C.; Puig, S.; et al. Validation of Artificial Intelligence Prediction Models for Skin Cancer Diagnosis Using Dermoscopy Images: The 2019 International Skin Imaging Collaboration Grand Challenge.Lancet Digit. Health2022,4, e330–e339

    2019

  5. [5]

    Shortcut Learning in Deep Neural Networks.Nat

    Geirhos, R.; Jacobsen, J.-H.; Michaelis, C.; Zemel, R.; Brendel, W.; Bethge, M.; Wichmann, F.A. Shortcut Learning in Deep Neural Networks.Nat. Mach. Intell.2020,2, 665–673

    2020

  6. [6]

    Uncovering and Correcting Shortcut Learning in Machine Learning Models for Skin Cancer Diagnosis.Diagnostics2022,12, 40

    Nauta, M.; Walsh, R.; Dubowski, A.; Seifert, C. Uncovering and Correcting Shortcut Learning in Machine Learning Models for Skin Cancer Diagnosis.Diagnostics2022,12, 40

  7. [7]

    A Survey on Image Data Augmentation for Deep Learning.J

    Shorten, C.; Khoshgoftaar, T.M. A Survey on Image Data Augmentation for Deep Learning.J. Big Data2019,6, 60

  8. [8]

    Gutman, D.; Codella, N.C.F.; Celebi, M.E.; Helba, B.; Marchetti, M.A.; Mishra, N.; Halpern, A. Skin Lesion Analysis Toward Melanoma Detection: A Challenge at the International Symposium on Biomedical Imaging (ISBI) 2016, Hosted by the International Skin Imaging Collaboration (ISIC).arXiv2016, arXiv:1605.01397

    Pith/arXiv arXiv 2016

  9. [9]

    Codella, N.C.F.; Gutman, D.; Celebi, M.E.; Helba, B.; Marchetti, M.A.; Dusza, S.W.; Kalloo, A.; Liopyris, K.; Mishra, N.; Kittler, H.; et al. Skin Lesion Analysis Toward Melanoma Detection: A Challenge at the 2017 International Symposium on Biomedical 27 of 27 Imaging (ISBI), Hosted by the International Skin Imaging Collaboration (ISIC). InProceedings of ...

    2017

  10. [10]

    Skin Lesion Analysis Toward Melanoma Detection 2018: A Challenge Hosted by the International Skin Imaging Collaboration (ISIC).arXiv2019, arXiv:1902.03368

    Codella, N.; Rotemberg, V .; Tschandl, P .; Celebi, M.E.; Dusza, S.; Gutman, D.; Helba, B.; Kalloo, A.; Liopyris, K.; Marchetti, M.; et al. Skin Lesion Analysis Toward Melanoma Detection 2018: A Challenge Hosted by the International Skin Imaging Collaboration (ISIC).arXiv2019, arXiv:1902.03368

    Pith/arXiv arXiv 2018

  11. [11]

    The HAM10000 Dataset, a Large Collection of Multi-Source Dermatoscopic Images of Common Pigmented Skin Lesions.Sci

    Tschandl, P .; Rosendahl, C.; Kittler, H. The HAM10000 Dataset, a Large Collection of Multi-Source Dermatoscopic Images of Common Pigmented Skin Lesions.Sci. Data2018,5, 180161

  12. [12]

    Seven-Point Checklist and Skin Lesion Classification Using Multitask Multimodal Neural Nets.IEEE J

    Kawahara, J.; Daneshvar, S.; Argenziano, G.; Hamarneh, G. Seven-Point Checklist and Skin Lesion Classification Using Multitask Multimodal Neural Nets.IEEE J. Biomed. Health Inform.2019,23, 538–546

    2019

  13. [13]

    Analysis of the ISIC Image Datasets: Usage, Benchmarks and Recommendations.Med

    Cassidy, B.; Kendrick, C.; Brodzicki, A.; Jaworek-Korjakowska, J.; Yap, M.H. Analysis of the ISIC Image Datasets: Usage, Benchmarks and Recommendations.Med. Image Anal.2022,75, 102305

    2022

  14. [14]

    Data Augmentation for Skin Lesion Analysis

    Perez, F.; Vasconcelos, C.; Avila, S.; Valle, E. Data Augmentation for Skin Lesion Analysis. InOR 2.0 Context-Aware Operating Theaters, Computer Assisted Robotic Endoscopy, Clinical Image-Based Procedures, and Skin Image Analysis (MICCAI 2018 Workshops); Lecture Notes in Computer Science, Vol. 11041; Springer: Cham, Switzerland, 2018; pp. 303–311

    2018

  15. [15]

    Data, Depth, and Design: Learning Reliable Models for Skin Lesion Analysis.Neurocomputing2020,383, 303–313

    Valle, E.; Fornaciali, M.; Menegola, A.; Tavares, J.; Bittencourt, F.V .; Li, L.T.; Avila, S. Data, Depth, and Design: Learning Reliable Models for Skin Lesion Analysis.Neurocomputing2020,383, 303–313

  16. [16]

    RandAugment: Practical Automated Data Augmentation with a Reduced Search Space

    Cubuk, E.D.; Zoph, B.; Shlens, J.; Le, Q.V . RandAugment: Practical Automated Data Augmentation with a Reduced Search Space. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW); IEEE: Piscataway, NJ, USA, 2020; pp. 702–703

    2020

  17. [17]

    mixup: Beyond Empirical Risk Minimization

    Zhang, H.; Cissé, M.; Dauphin, Y.N.; Lopez-Paz, D. mixup: Beyond Empirical Risk Minimization. InProceedings of the International Conference on Learning Representations (ICLR), Vancouver, BC, Canada, 30 April–3 May 2018

    2018

  18. [18]

    (De)Constructing Bias on Skin Lesion Datasets

    Bissoto, A.; Fornaciali, M.; Valle, E.; Avila, S. (De)Constructing Bias on Skin Lesion Datasets. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW); IEEE: Piscataway, NJ, USA, 2019; pp. 2766–2774

    2019

  19. [19]

    Disparities in Dermatology AI Performance on a Diverse, Curated Clinical Image Set.Sci

    Daneshjou, R.; Vodrahalli, K.; Novoa, R.A.; Jenkins, M.; Liang, W.; Rotemberg, V .; Ko, J.; Swetter, S.M.; Bailey, E.E.; Gevaert, O.; et al. Disparities in Dermatology AI Performance on a Diverse, Curated Clinical Image Set.Sci. Adv.2022,8, eabq6147

    2022

  20. [20]

    Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fitzpatrick 17k Dataset

    Groh, M.; Harris, C.; Soenksen, L.; Lau, F.; Han, R.; Kim, A.; Koochek, A.; Badri, O. Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fitzpatrick 17k Dataset. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW); IEEE: Piscataway, NJ, USA, 2021; pp. 1820–1828

    2021

  21. [21]

    Domain-Adversarial Training of Neural Networks.J

    Ganin, Y.; Ustinova, E.; Ajakan, H.; Germain, P .; Larochelle, H.; Laviolette, F.; Marchand, M.; Lempitsky, V . Domain-Adversarial Training of Neural Networks.J. Mach. Learn. Res.2016,17, 1–35

    2016

  22. [22]

    Tent: Fully Test-Time Adaptation by Entropy Minimization

    Wang, D.; Shelhamer, E.; Liu, S.; Olshausen, B.; Darrell, T. Tent: Fully Test-Time Adaptation by Entropy Minimization. In Proceedings of the International Conference on Learning Representations (ICLR), Virtual, 3–7 May 2021

    2021

  23. [23]

    Visualizing Data Using t-SNE.J

    van der Maaten, L.; Hinton, G. Visualizing Data Using t-SNE.J. Mach. Learn. Res.2008,9, 2579–2605

    2008

  24. [24]

    A ConvNet for the 2020s

    Liu, Z.; Mao, H.; Wu, C.-Y.; Feichtenhofer, C.; Darrell, T.; Xie, S. A ConvNet for the 2020s. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE: Piscataway, NJ, USA, 2022; pp. 11976–11986

    2022

  25. [25]

    A Method for Normalizing Histology Slides for Quantitative Analysis

    Macenko, M.; Niethammer, M.; Marron, J.S.; Borland, D.; Woosley, J.T.; Guan, X.; Schmitt, C.; Thomas, N.E. A Method for Normalizing Histology Slides for Quantitative Analysis. InProceedings of the 2009 IEEE International Symposium on Biomedical Imaging (ISBI); IEEE: Piscataway, NJ, USA, 2009; pp. 1107–1110

    2009

  26. [26]

    Quantifying the Effects of Data Augmentation and Stain Color Normalization in Convolutional Neural Networks for Computational Pathology.Med

    Tellez, D.; Litjens, G.; Bándi, P .; Bulten, W.; Bokhorst, J.-M.; Ciompi, F.; van der Laak, J. Quantifying the Effects of Data Augmentation and Stain Color Normalization in Convolutional Neural Networks for Computational Pathology.Med. Image Anal. 2019,58, 101544

    2019

  27. [27]

    AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty

    Hendrycks, D.; Mu, N.; Cubuk, E.D.; Zoph, B.; Gilmer, J.; Lakshminarayanan, B. AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty. InProceedings of the International Conference on Learning Representations (ICLR), Addis Ababa, Ethiopia, 26–30 April 2020

    2020

  28. [28]

    Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization

    Selvaraju, R.R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; Batra, D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. InProceedings of the IEEE International Conference on Computer Vision (ICCV); IEEE: Piscataway, NJ, USA, 2017; pp. 618–626

    2017

  29. [29]

    In Search of Lost Domain Generalization

    Gulrajani, I.; Lopez-Paz, D. In Search of Lost Domain Generalization. InProceedings of the International Conference on Learning Representations (ICLR), Virtual, 3–7 May 2021

    2021

  30. [30]

    A Simple Framework for Contrastive Learning of Visual Representations

    Chen, T.; Kornblith, S.; Norouzi, M.; Hinton, G. A Simple Framework for Contrastive Learning of Visual Representations. In Proceedings of the 37th International Conference on Machine Learning (ICML); PMLR: Cambridge, MA, USA, 2020; pp. 1597–1607

    2020

  31. [31]

    Test-Time Training with Self-Supervision for Generalization under Distribution Shifts

    Sun, Y.; Wang, X.; Liu, Z.; Miller, J.; Efros, A.A.; Hardt, M. Test-Time Training with Self-Supervision for Generalization under Distribution Shifts. InProceedings of the 37th International Conference on Machine Learning (ICML); PMLR: Cambridge, MA, USA, 2020; pp. 9229–9248

    2020