Pith. sign in

REVIEW 4 major objections 6 minor 44 references

Safer Skin Lesion Classification with Global Class Activation Probability Map Evaluation and SafeML

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read All-class activation maps, measured against lesion ground truth, can act as runtime safety indicators that flag unreliable skin-lesion predictions before they are issued.

desk verdict The multi-class attribution idea is worth a look, but Eq. (2)'s undefined probability makes the reported numbers non-reproducible as written. read the letter →

arxiv 2508.20776 v1 pith:SEHM7VOD submitted 2025-08-28 cs.CV cs.AI

Kuniko Paxton , Koorosh Aslansefat , Amila Akagić , Dhavalkumar Thakker , Yiannis Papadopoulos This is my paper
classification cs.CVcs.AI
keywords skinlesionclassificationexplainabilityclassactivationmappingGrad-CAMattributesensitivityfalsepositiverateselectivepredictionSafeML
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 aims to turn explainability from a passive visualization into an active safety check for skin-lesion classifiers. It introduces GCAPM, which computes class activation maps for every candidate class and labels each image pixel with the class the model most strongly activates, yielding a segmented attention map that can be quantitatively compared to lesion annotations. Two metrics, Att Sensitivity and Att FPR, measure how much of the lesion and how much non-lesion area the model attended to; the paper reports that these metrics correlate with prediction correctness and can be used to build a SafeML-inspired selective predictor. With MobileNetV2 and Vision Transformers on ISIC 2017 and ISIC 2019, the predictor accepted correct diagnoses with roughly 90% accuracy and flagged incorrect ones with over 75% accuracy under simulated image degradation. If the claim holds, the same model that produces a diagnosis can also decide when that diagnosis should not be trusted without waiting for a label.

What carries the argument

The load-bearing object is GCAPM, the Global Class Activation Probabilistic Map: for every class c, a Grad-CAM map over the input image is computed and treated as a per-pixel probability-like score, so Eq. (2) can select, pixel by pixel, the class with maximum P(c|cam(h,w)). The result is a single segmentation-like image in which each pixel is labeled by the class the model most strongly associates with it. Two spatial overlap metrics, Att Sensitivity and Att FPR, compare that label map to the ground-truth lesion mask. The second mechanism is a SafeML-style selective predictor: a binary SVM meta-classifier that takes the lesion model's predicted probabilities together with the two attribute

What would settle it

Take a trained classifier and compute GCAPM attribute metrics after randomly permuting the class assignment of each pixel's activation map, so the spatial maps no longer correspond to genuine class attention. If a meta-classifier trained on the permuted metrics still flags incorrect predictions with over 75% accuracy, then the safety signal is not coming from the claimed class-specific explanation content, and the central claim is undermined.

Watch

Extended reading notes

Core claim

The paper's central claim is that explainability can be made quantitative and safety-relevant by looking at every class at once. Its GCAPM method computes a Grad-CAM activation map for each candidate class, normalizes these maps, and assigns each pixel to the class with the highest activation, producing a class-labeled segmentation of the model's attention. Comparing that segmentation output to the lesion mask yields two attribute metrics: Att Sensitivity, how much of the lesion area the model attended to, and Att FPR, how much non-lesion area it attended to. Across MobileNetV2 and Vision Transformer classifiers on ISIC 2017 and ISIC 2019, inaccurate predictions clustered at low Att Sensitiv

Load-bearing premise

The approach assumes that normalizing each class's Grad-CAM map and taking a pixel-wise maximum yields a meaningful probability that a pixel belongs to a class, so the resulting attribute-metric scores genuinely measure the model's diagnostic focus.

Editorial extensions

If this is right

  • Skin lesion classifiers can be evaluated on whether they look at the lesion region across all candidate classes, not only the predicted one, making single-class explanations harder to over-trust.
  • Attribute sensitivity and attribute FPR can be used as quantitative trust indicators: low sensitivity and high false-positive attention flag a prediction for human review.
  • A selective predictor trained on predicted probabilities plus attribute metrics can operate at runtime without ground-truth labels and trigger human intervention on suspect diagnoses.
  • Under 10% to 50% Gaussian blur, the selective predictor kept roughly 90% accuracy on identifying correct predictions and flagged over 75% of incorrect ones in most tested settings.
  • Predictions inside the attribute-metric confidence interval were more often correct than predictions with softmax confidence above 50%, suggesting explainability-derived confidence is a stronger acceptance criterion than raw probability.

Reading between the lines

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

  • The paper does not test the method on black-box APIs; a model-agnostic attribution step would be needed to extend the same safety logic to commercial systems.
  • The paper simulates drift only with Gaussian blur; testing with a different acquisition device or skin-tone distribution would show whether the confidence interval generalizes.
  • The paper does not ablate the pixel-wise normalization in Eq. (2); comparing it against class-output-weighted Grad-CAM would isolate whether the safety signal comes from spatial attention or class weights.
  • The same Att-metrics design could transfer to other medical imaging tasks that have binary region-of-interest masks, since it requires only lesion/background labels.
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 / 6 minor

Summary. The paper proposes GCAPM, a method that assigns each pixel to the class with the highest per-class Grad-CAM-derived activation probability, visualizes this across all classes, and derives two attribute metrics (Att Sensitivity, Att FPR) by comparing the resulting segmentation to lesion ground truth. It then uses these metrics, together with prediction probabilities, as inputs to an SVM-based selective predictor that is meant to flag potentially incorrect diagnoses at runtime. The method is evaluated on ISIC 2017 and ISIC 2019 using MobileNetV2 and Vision Transformer classifiers, including blurred versions of the data to simulate drift. The paper claims that GCAPM improves explainability reliability, that attribute metrics correlate with prediction performance, and that the selective predictor detects incorrect predictions with high accuracy.

Significance. If the central claims held, the paper would make a useful contribution: it extends CAM-based explanation from a single predicted class to a global, all-class pixel-level map, and it couples that map with a SafeML-style selective prediction mechanism, potentially improving human oversight in automated skin lesion screening. The availability of code, the use of two architectures and two benchmark datasets, and the external check on blurred runtime data are strengths. However, the central quantitative claims are not currently supported: the definition of the GCAPM probability map is incomplete, RQ1 is not answered by any comparison against conventional CAM methods, and one of the four model/dataset combinations shows a negative correlation that contradicts the paper's reliability claim. The overall idea is promising and within the scope of the venue, but the load-bearing gaps described below need to be addressed before the claims can be accepted.

major comments (4)
  1. [§3.1, Eq. (2)] Eq. (2) defines C_{h,w} = argmax_c P(c|cam(h,w)), but P(c|cam(h,w)) is never defined. The text only says that class activation maps are 'normalizing and probabilistically aggregating at a pixel level.' Per-class Grad-CAM maps have class-dependent and scale-arbitrary values (they are ReLU outputs of differently weighted gradient combinations), so an argmax over classes is meaningless without a precise normalization. This ambiguity is inherited by every downstream quantity: Att Sensitivity (Eq. 3), Att FPR (Eq. 4), the confidence intervals in Fig. 3, and the SVM inputs in Eq. (5). Without a reproducible definition of P, the reported quantitative results in Tables 1 and 2 cannot be independently verified. Please specify the exact normalization (e.g., per-pixel softmax over class maps after some scaling) and justify why the resulting values represent a meaningful class probability.
  2. [§2.3, RQ1 and §5.1] RQ1 asks whether GCAPM 'improves the reliability of explainability ... compared to conventional CAM-based methods,' but no quantitative comparison to any baseline CAM method is provided. Fig. 2 is illustrative only, and the attribute metrics in Tables 1-2 are computed only for GCAPM. To answer RQ1, the paper needs a comparison of GCAPM against, at minimum, standard Grad-CAM (single-class) and possibly a class-wise Grad-CAM baseline, using the same attribute metrics or established explanation-quality measures (e.g., localization accuracy, IoU with ground-truth lesion). As written, the claim that GCAPM improves reliability over conventional CAM is not supported.
  3. [Table 1, §5.2] The paper states that Att Sensitivity and Att FPR 'can serve as supplementary indicators of trust,' but Table 1 shows a negative correlation for ISIC 2019 ViT: Att Sensitivity vs. F1 score is -0.14 and vs. accuracy is -0.12, with Att FPR essentially 0.00/0.03. This is the opposite of the claimed relationship and directly undermines the generalization of the reliability claim. Even the positive correlations are modest (e.g., 0.28/0.35, 0.15/0.22), so the conclusion that these metrics are reliable indicators needs to be qualified by architecture and dataset. Please either provide an explanation for the negative correlation, report the result honestly as a limitation, or revise the claim.
  4. [§3.3, §5.3, Table 2] The selective prediction framework is central to RQ3, but the training and evaluation protocol is underspecified. How is the SVM meta-classifier trained? Is it fitted on offline ISIC data only, and if so, exactly which samples are used for fitting the confidence intervals and the SVM? Table 2 reports columns 'Acc' and 'Inacc' without defining them clearly (presumably the selective predictor's accuracy on correct and incorrect predictions, respectively), and no comparison is made to a baseline that selects by prediction probability alone, even though Fig. 5 implicitly contrasts them. Without this information, it is difficult to assess whether the reported 90%/75% figures reflect the method or the evaluation setup, and whether the good performance is a result of fitting the SVM and thresholds to the same offline distribution that is then blurred.
minor comments (6)
  1. [§3.1, Eq. (1)] The text says 'as shown in Eq. 1' twice in the same paragraph, and the second sentence refers to Eq. 1 again when Eq. (2) is meant. Please fix the equation references.
  2. [§5.3 / Table 2] The table headers 'Acc' and 'Inacc' are ambiguous. Please clarify whether these are the selective predictor's classification accuracies for predictions that were actually correct/incorrect, and add a caption explaining the entries.
  3. [§4.1] The generation of segmentation masks for ISIC 2019 via DeepLabV3 fine-tuned on HAM10000 is described in one sentence. Since the attribute metrics depend on these masks, please provide details on preprocessing, validation, and any filtering of low-confidence masks.
  4. [§5.2 / Fig. 3] The axes of Fig. 3 are labeled in an abbreviated way, and the caption does not clearly describe what is plotted (density of which variable versus what?). Please make the figure self-contained.
  5. [§6.3] The limitations paragraph correctly notes that ViT interpretability is still immature and that the method requires internal model access, but the conclusion still claims the system 'contributes to enabling safer and more reliable use of DL in medical settings.' Please connect the limitations more explicitly to the conclusions, for instance by stating that the safety guarantee does not apply to black-box commercial systems.
  6. [Throughout] There are several typos: 'interoperability' should likely be 'interpretability' in §2.1; 'idendifed' and 'ot' in §5.3; 'explanability' in the conclusion. A careful proofread is needed.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation; the paper's safety metrics are empirical evaluations against external annotations and blurred runtime data. The undefined P(c|cam) in Eq. (2) is a reproducibility gap, not a circular reduction.

full rationale

Walking the derivation chain, I find no step where a claimed prediction or first-principles result is equivalent to its inputs by construction. The per-pixel class assignment in Eq. (2) is defined as argmax_c P(c|cam(h,w)); although P(c|cam) is never explicitly defined (Sec. 3.1, Algorithm 1), this is a missing definition and reproducibility gap rather than a circular reduction: the subsequent Att Sensitivity (Eq. 3) and Att FPR (Eq. 4) are compared against external ISIC lesion annotations (or DeepLabV3-generated proxy masks for ISIC 2019), and the selective predictor in Sec. 5.3 is evaluated on blurred runtime data that the offline thresholds/SVM are stated to be established from during offline development, making the runtime evaluation an external check. The SafeML citations [7,5,3] are from the current research group, but they are used as an existing statistical drift-monitoring concept, not as a uniqueness theorem or as a substitute for the paper's own experiments. The only load-bearing gap is the undefined P(c|cam), which prevents independent reproduction but does not make the output equal to the input by construction. No circularity score above 1 is warranted.

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

The method relies on an unspecified probability normalization, fitted CI thresholds, and a fitted SVM, plus several domain assumptions about the datasets and segmentation masks. There are no newly invented physical or conceptual entities beyond the GCAPM map itself.

free parameters (3)
  • P(c|cam) normalization = not reported
    Eq. 2 requires converting per-class CAM values to probabilities; if a softmax with temperature or scaling is used, the temperature or scale is a free parameter that is never specified.
  • Attribute metric confidence interval thresholds = not reported
    The CI used in Fig. 5 and Table 2 is derived from offline distributions of Att Sensitivity and Att FPR, but the exact quantiles or coverage values are not given.
  • SVM hyperparameters = not reported
    The selective predictor uses an SVM, but the kernel, C, gamma, and other hyperparameters are not stated, and no train/test split or cross-validation is described.
assumptions (4)
  • ad hoc to paper Pixel-wise argmax over class CAM values yields a meaningful class-association map.
    Eq. 2 assumes this without justification; the relationship between CAM magnitude and class probability is not established.
  • domain assumption The three selected classes from ISIC 2019 are representative and match the lesion segmentation task.
    ISIC 2019 originally has eight classes, but the paper reduces to three without naming them, making the comparison to lesion masks ambiguous.
  • domain assumption DeepLabV3-generated masks for ISIC 2019 are a valid proxy for expert lesion annotations.
    The paper uses generated masks (Mean IoU 88%) as ground truth for Att metric computation, which introduces potential error not accounted for in the results.
  • standard math Grad-CAM is correctly computed for both MobileNetV2 and Vision Transformer.
    Grad-CAM is a standard method, but its application to ViT is known to be less interpretable, which the paper acknowledges in limitations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Safer Skin Lesion Classification with Global Class Activation Probability Map Evaluation and SafeML." pith.science (2026). https://pith.science/paper/SEHM7VOD

@misc{pith2026250820776,
  author       = {Pith},
  title        = {Pith review of: Safer Skin Lesion Classification with Global Class Activation Probability Map Evaluation and SafeML},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SEHM7VOD}},
  note         = {Machine review of arXiv:2508.20776}
}
read the original abstract

Recent advancements in skin lesion classification models have significantly improved accuracy, with some models even surpassing dermatologists' diagnostic performance. However, in medical practice, distrust in AI models remains a challenge. Beyond high accuracy, trustworthy, explainable diagnoses are essential. Existing explainability methods have reliability issues, with LIME-based methods suffering from inconsistency, while CAM-based methods failing to consider all classes. To address these limitations, we propose Global Class Activation Probabilistic Map Evaluation, a method that analyses all classes' activation probability maps probabilistically and at a pixel level. By visualizing the diagnostic process in a unified manner, it helps reduce the risk of misdiagnosis. Furthermore, the application of SafeML enhances the detection of false diagnoses and issues warnings to doctors and patients as needed, improving diagnostic reliability and ultimately patient safety. We evaluated our method using the ISIC datasets with MobileNetV2 and Vision Transformers.

Figures

Figures reproduced from arXiv: 2508.20776 by the authors.

Figure 1
Figure 1. Safety Skin Lesion Detection Procedure: Using the knowledge gained from of [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Risk Case 1 is Grad-CAM in classes of ISIC 2019. The red line shows the [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Density of Attributes performance: (A) to (D) correspond to the combinations of [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Attributes and Predictive Performance each Data: (A) shows the MobileNet [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Prediction Accuracy in CI, Out of CI, and Probability [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 27 canonical work pages

  1. [1]

    arXiv preprint arXiv:2005.00928 (2020)

    Abnar, S., Zuidema, W.: Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928 (2020)

    arXiv 2005

  2. [2]

    Nature Machine Intelligence5(9), 1006– 1019 (2023)

    Achtibat, R., Dreyer, M., Eisenbraun, I., Bosse, S., Wiegand, T., Samek, W., Lapuschkin, S.: From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence5(9), 1006– 1019 (2023)

    2023

  3. [3]

    In: Computer Safety, Reliability, and Security

    Akram, M.N., Ambekar, A., Sorokos, I., Aslansefat, K., Schneider, D.: Stadre and stadro: Reliability and robustness estimation of ml-based forecasting using statisti- cal distance measures. In: Computer Safety, Reliability, and Security. SAFECOMP 2022 Workshops: DECSoS, DepDevOps, SASSUR, SENSEI, USDAI, and WAISE Munich, Germany, September 6–9, 2022, Proc...

    2022

  4. [4]

    IEEE Software (2023)

    Aslansefat, K., Hashemian, M., Walker, M., Akram, M.N., Sorokos, I., Papadopou- los, Y.: Explaining black boxes with a smile: Statistical model-agnostic inter- pretability with local explanations. IEEE Software (2023)

    2023

  5. [5]

    Computer54(8), 66–76 (2021)

    Aslansefat, K., Kabir, S., Abdullatif, A., Vasudevan, V., Papadopoulos, Y.: To- ward improving confidence in autonomous vehicle software: A study on traffic sign recognition systems. Computer54(8), 66–76 (2021)

    2021

  6. [6]

    In: Model-Based Safety and Assessment: 8th International Symposium, IMBSA 2022, Munich, Germany, September 5–7, 2022, Proceedings

    Aslansefat, K., Nikolaou, P., Walker, M., Akram, M.N., Sorokos, I., Reich, J., Kolios, P., Michael, M.K., Theocharides, T., Ellinas, G., et al.: Safedrones: Real- time reliability evaluation of uavs using executable digital dependable identities. In: Model-Based Safety and Assessment: 8th International Symposium, IMBSA 2022, Munich, Germany, September 5–7...

    2022

  7. [7]

    In: International symposium on model-based safety and assessment

    Aslansefat, K., Sorokos, I., Whiting, D., Tavakoli Kolagari, R., Papadopoulos, Y.: Safeml: safety monitoring of machine learning classifiers through statistical differ- ence measures. In: International symposium on model-based safety and assessment. pp. 197–211. Springer (2020)

    work page 2020

  8. [8]

    Bergler, M., Kolagari, R.T., Lundqvist, K.: Case study on the use of the safeml ap- proachintrainingautonomousdrivingvehicles.In:ImageAnalysisandProcessing– ICIAP2022:21stInternationalConference,Lecce,Italy,May23–27,2022,Proceed- ings, Part III. pp. 87–97. Springer (2022)

    2022

Show all 44 references
  1. [9]

    arXiv preprint arXiv:2307.04137 (2023)

    Cao, Q.H., Nguyen, T.T.H., Nguyen, V.T.K., Nguyen, X.P.: A novel explain- able artificial intelligence model in image classification problem. arXiv preprint arXiv:2307.04137 (2023)

    2023 arXiv

  2. [10]

    Medical Image Analysis (2021)

    Cassidy, B., Kendrick, C., Brodzicki, A., Jaworek-Korjakowska, J., Yap, M.H.: Analysis of the isic image datasets: Usage, benchmarks and recommendations. Medical Image Analysis (2021)

    2021

  3. [11]

    Neurocomputing408, 189–215 (2020)

    Cervantes, J., Garcia-Lamont, F., Rodríguez-Mazahua, L., Lopez, A.: A compre- hensive survey on support vector machine classification: Applications, challenges and trends. Neurocomputing408, 189–215 (2020)

    2020

  4. [12]

    Nature Communications15(1), 524 (2024)

    Chanda, T., Hauser, K., Hobelsberger, S., Bucher, T.C., Garcia, C.N., Wies, C., Kittler, H., Tschandl, P., Navarrete-Dechent, C., Podlipnik, S., et al.: Dermatologist-like explainable ai enhances trust and confidence in diagnosing melanoma. Nature Communications15(1), 524 (2024)

    2024

  5. [13]

    ACM trans- actions on intelligent systems and technology (TIST)2(3), 1–27 (2011)

    Chang, C.C., Lin, C.J.: Libsvm: a library for support vector machines. ACM trans- actions on intelligent systems and technology (TIST)2(3), 1–27 (2011)

    2011

  6. [14]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chefer, H., Gur, S., Wolf, L.: Transformer interpretability beyond attention visu- alization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 782–791 (2021) 14 K.Paxton et al

    2021

  7. [15]

    IEEE Access10, 39769–39781 (2022)

    Cho, H., Lee, K., Choi, N., Kim, S., Lee, J., Yang, S.: Online safety zone estimation and violation detection for nonstationary objects in workplaces. IEEE Access10, 39769–39781 (2022)

    2022

  8. [16]

    In: 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018)

    Codella, N.C., 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 biomed- ical imaging (isbi), hosted b...

    2017

  9. [17]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    2010 arXiv

  10. [18]

    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)

    2017

  11. [19]

    Machine Learning Research9, 1871–1874 (2008)

    Fan, R.E., Chang, K.W., Hsieh, C.J., Wang, X.R., Lin, C.J.: Liblinear: A library for large linear classification. Machine Learning Research9, 1871–1874 (2008)

    2008

  12. [20]

    arXiv preprint arXiv:2312.10801 (2023)

    Farhad, A.H., Sorokos, I., Akram, M.N., Aslansefat, K., Schneider, D.: Scope com- pliance uncertainty estimate. arXiv preprint arXiv:2312.10801 (2023)

    2023 arXiv

  13. [21]

    In: Model-Based Safety and Assessment: 8th International Symposium, IMBSA 2022, Munich, Germany, September 5–7, 2022, Proceedings

    Farhad, A.H., Sorokos, I., Schmidt, A., Akram, M.N., Aslansefat, K., Schneider, D.: Keep your distance: Determining sampling and distance thresholds in machine learning monitoring. In: Model-Based Safety and Assessment: 8th International Symposium, IMBSA 2022, Munich, Germany,...

    2022

  14. [22]

    Gertych, A., Faust, O.: Ai explainability and bias propagation in medical decision support (2024)

    2024

  15. [23]

    European Journal of Cancer 167, 54–69 (2022)

    Hauser, K., Kurz, A., Haggenmüller, S., Maron, R.C., von Kalle, C., Utikal, J.S., Meier, F., Hobelsberger, S., Gellrich, F.F., Sergon, M., et al.: Explainable artificial intelligence in skin cancer recognition: A systematic review. European Journal of Cancer 167, 54–69 (2022)

    2022

  16. [24]

    Scientific data11(1), 641 (2024)

    Hernández-Pérez, C., Combalia, M., Podlipnik, S., Codella, N.C., Rotemberg, V., Halpern, A.C., Reiter, O., Carrera, C., Barreiro, A., Helba, B., et al.: Bcn20000: Dermoscopic lesions in the wild. Scientific data11(1), 641 (2024)

    2024

  17. [25]

    Applied Soft Computing159, 111624 (2024)

    Hosny, K.M., Said, W., Elmezain, M., Kassem, M.A.: Explainable deep inherent learning for multi-classes skin lesion classification. Applied Soft Computing159, 111624 (2024)

    2024

  18. [26]

    In: World Conference on Explainable Artificial Intelligence

    Hryniewska-Guzik, W., Longo, L., Biecek, P.: Cnn-based explanation ensembling for dataset, representation and explanations evaluation. In: World Conference on Explainable Artificial Intelligence. pp. 346–368. Springer (2024)

    2024

  19. [27]

    In: 2022 International Conference on Computing, Electronics & Communications Engineering (iCCECE)

    Kabir, S., Aslansefat, K., Gope, P., Campean, F., Papadopoulos, Y.: Combining drone-based monitoring and machine learning for online reliability evaluation of wind turbines. In: 2022 International Conference on Computing, Electronics & Communications Engineering (iCCECE). pp. ...

    2022

  20. [28]

    Kläs, M.: DIN SPEC 92005: Machine Learning – Uncertainty Quantification. DIN SPEC 92005, Beuth Verlag GmbH (DIN - Deutsches Institut für Normung), Berlin, Germany (2022), https://www.din.de/en/wdc-beuth:din21:343195966, standard for methods and measures for quantifying uncerta...

    2022

  21. [29]

    arXiv preprint arXiv:2111.11863 (2021) Title Suppressed Due to Excessive Length 15

    Metta, C., Beretta, A., Guidotti, R., Yin, Y., Gallinari, P., Rinzivillo, S., Gian- notti, F.: Explainable deep image classifiers for skin lesion diagnosis. arXiv preprint arXiv:2111.11863 (2021) Title Suppressed Due to Excessive Length 15

    2021 arXiv

  22. [30]

    Nigar, N., Umar, M., Shahzad, M.K., Islam, S., Abalo, D.: A deep learning ap- proachbasedonexplainableartificialintelligenceforskinlesionclassification.IEEE Access 10, 113715–113725 (2022)

    2022

  23. [31]

    In: International cross- domain conference for machine learning and knowledge extraction

    Nunnari, F., Kadir, M.A., Sonntag, D.: On the overlap between grad-cam saliency maps and explainable visual features in skin cancer images. In: International cross- domain conference for machine learning and knowledge extraction. pp. 241–253. Springer (2021)

    2021

  24. [32]

    Neural Computing and Applications33(22), 15171–15189 (2021)

    Pintelas, E., Liaskos, M., Livieris, I.E., Kotsiantis, S., Pintelas, P.: A novel explain- able image classification framework: case study on skin cancer and plant disease prediction. Neural Computing and Applications33(22), 15171–15189 (2021)

    2021

  25. [33]

    why should i trust you?

    Ribeiro, M.T., Singh, S., Guestrin, C.: " why should i trust you?" explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD. pp. 1135– 1144 (2016)

    2016

  26. [34]

    Neurocomputing597, 127969 (2024)

    Saadatfar, H., Kiani-Zadegan, Z., Ghahremani-Nezhad, B.: Us-lime: Increasing fi- delity in lime using uncertainty sampling on tabular data. Neurocomputing597, 127969 (2024)

    2024

  27. [35]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: In- verted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510–4520 (2018)

    2018

  28. [36]

    In: Proceedings of the IEEE Int

    Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad- cam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE Int. Conf. on computer vision. pp. 618–626 (2017)

    2017

  29. [37]

    Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: Int. Conf. on machine learning. pp. 3319–3328. PMLR (2017)

    2017

  30. [38]

    Nature Medicine26(8), 1229–1234 (2020)

    Tschandl, P., Rinner, C., Apalla, Z., Argenziano, G., Codella, N., Halpern, A., Janda, M., Lallas, A., Longo, C., Malvehy, J., et al.: Human–computer collabora- tion for skin cancer recognition. Nature Medicine26(8), 1229–1234 (2020)

    2020

  31. [39]

    Scientific data 5(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 data 5(1), 1–9 (2018)

    2018

  32. [40]

    In: Journal of Physics: Conference Series

    Walker, C., Rothon, C., Aslansefat, K., Papadopoulos, Y., Dethlefs, N.: Using large language models to recommend repair actions for offshore wind maintenance. In: Journal of Physics: Conference Series. vol. 2875, p. 012025. IOP Publishing (2024)

    2024

  33. [41]

    In: Proceedings of the IEEE/CVF conference on computer vision and pat- tern recognition workshops

    Wang, H., Wang, Z., Du, M., Yang, F., Zhang, Z., Ding, S., Mardziel, P., Hu, X.: Score-cam: Score-weighted visual explanations for convolutional neural net- works. In: Proceedings of the IEEE/CVF conference on computer vision and pat- tern recognition workshops. pp. 24–25 (2020)

    2020

  34. [42]

    Applied Intelligence 53(23), 28226–28240 (2023)

    Xiang, X., Yu, H., Wang, Y., Wang, G.: Stable local interpretable model-agnostic explanations based on a variational autoencoder. Applied Intelligence 53(23), 28226–28240 (2023)

    2023

  35. [43]

    arXiv preprint arXiv:2101.08418 (2021)

    Zhang, Y., Mehta, S., Caspi, A.: Rethinking semantic segmentation evaluation for explainability and model selection. arXiv preprint arXiv:2101.08418 (2021)

    2021 arXiv

  36. [44]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921–2929 (2016)

    2016

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.