Pith. sign in

REVIEW 3 major objections 5 minor 8 references

Using Computer Vision for Skin Disease Diagnosis in Bangladesh Enhancing Interpretability and Transparency in Deep Learning Models for Skin Cancer Classification

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read The paper claims that a custom convolutional network can classify skin-lesion images with high accuracy and can show clinicians the image regions behind each decision.

desk verdict The paper promises saliency and attention maps but never produces them; the accuracy claim is also undermined by test-set filtering, so neither half of the central claim survives. read the letter →

arxiv 2501.18161 v1 pith:OTCICNF3 submitted 2025-01-30 eess.IV cs.CV

classification eess.IVcs.CV
keywords skincancerclassificationdeepconvolutionalneuralnetworktransferlearningsaliencymapsattentioninterpretabilityHAM10000Bangladesh
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 show that a purpose-built deep convolutional neural network can classify dermoscopic skin-lesion images as benign or malignant with accuracy that matches or beats standard transfer-learning models, while also making its decisions inspectable through saliency and attention maps. The motivation is concrete: Bangladesh has few dermatologists, so an automated, explainable screening tool could move detection earlier. On the HAM10000 dataset, the proposed model reports testing accuracy around 90 percent, an AUC of 0.847, and lower per-epoch training time than AlexNet, ResNet, VGG-16, DenseNet, and MobileNet. The interpretability claim is that clinicians can look at the visualization maps and see which image regions drove the diagnosis.

What carries the argument

The load-bearing machinery is the custom DCNN pipeline: a preprocessing stage that removes reflective artifacts via a local-mean threshold, manually discards blurry and low-contrast images, normalizes and augments the data, and then a convolutional architecture with max pooling, binary cross-entropy loss, the Adam optimizer, and a final sigmoid layer. Interpretability comes from saliency maps, which mark pixel-level importance derived from the model's gradients, and attention maps, which show the regions the network focuses on; these are overlaid on the input images so a clinician can see where the model is looking. These maps are the mechanism claimed to add transparency without an accuracy penalty.

What would settle it

Re-run evaluation on the original unfiltered HAM10000 images, including the ones removed as blurry or artifact-heavy, and check whether accuracy and AUC stay near the reported values; they should if the filter was not responsible for the result.

Watch

Extended reading notes

Core claim

The central claim is that a custom DCNN, trained on HAM10000 after reflective-artifact removal, manual exclusion of blurry and low-contrast images, class balancing, and PCA-based augmentation, and using a sigmoid output for binary classification, outperforms five transfer-learning baselines on the same data. The authors report testing accuracy around 90 percent, an ROC AUC of 0.847, and faster per-epoch training than the baselines. They further argue that overlaying saliency maps and attention maps on input images highlights the pixels the model weighs most heavily, giving clinicians a visual explanation for each benign-versus-malignant decision, and claim this interpretability comes without materially reducing accuracy.

Load-bearing premise

The whole accuracy result depends on the assumption that deleting images because they look blurry, low-contrast, or artifact-heavy does not also delete a meaningful slice of the patient population; if it does, the reported accuracy does not transfer to new, uncurated images.

Editorial extensions

If this is right

  • If the reported results generalize, a binary DCNN classifier can serve as an automated second reader for benign-versus-malignant screening where dermatologists are scarce.
  • The saliency and attention maps give clinicians a concrete artifact to inspect: a diagnosis can be questioned when the highlighted region falls outside the lesion.
  • The runtime comparison implies the proposed model is lighter than the transfer-learning baselines, so screening could run on ordinary laptops or modest cloud instances.
  • The same preprocessing and augmentation recipe can be applied to other dermoscopic datasets, since the classifier itself is not tied to HAM10000.

Reading between the lines

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

  • The paper leaves open whether the saliency and attention maps align with what a dermatologist would consider the lesion; comparing map highlights to expert-annotated borders would quantify the interpretability claim.
  • Because the test set excluded visually imperfect images, the accuracy figures should not be read as a prediction of performance on uncurated smartphone images; that gap is an inference from the paper's own data-preparation section, not a claim it makes.
  • The favorable runtime comparison suggests the DCNN might run on modest hardware, but the paper does not measure inference time in a clinical setting, so real-world latency remains an open question.
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

3 major / 5 minor

Summary. The manuscript proposes a deep convolutional neural network (DCNN) for binary classification of benign versus malignant skin lesions from the HAM10000 dataset, compares it against AlexNet, ResNet, VGG-16, DenseNet, and MobileNet, and claims that the proposed model achieves the highest classification accuracy with lower execution time. The abstract and introduction further claim that the method enhances interpretability by combining saliency maps and attention maps to visualize features influencing model decisions. The body reports training/testing accuracy, precision, recall, F1-score, ROC AUC (0.847), and execution times for the proposed model and baselines, but it contains no saliency or attention maps, no procedure for generating them, and no interpretability evaluation.

Significance. If the advertised interpretability contribution were actually implemented and evaluated, the work could be relevant to computer-aided skin cancer diagnosis in settings with limited dermatologist access. Two aspects could have given the paper value: the explicit focus on a Bangladesh deployment scenario, and the effort to compare multiple transfer-learning baselines on the same data. However, as submitted, the paper does not deliver the interpretability method that motivates the title and abstract, and the accuracy claims are weakened by a non-representative test set and the absence of uncertainty quantification. The reported comparisons are therefore neither a valid generalization estimate nor evidence for the central interpretability claim.

major comments (3)
  1. [Abstract and Sections 1.2, 5-7] The central claim that the method 'enhance[s] the interpretability' of skin cancer classification through saliency and attention maps is unsupported by the manuscript body. The abstract and introduction promise a technique that visualizes critical features, but no equation, algorithm, layer description, or procedure for generating saliency or attention maps appears in Sections 4-7. Figures 6-9 show sample lesions, accuracy/loss curves, and an ROC curve only; no saliency or attention map is displayed, and no interpretability metric, localization benchmark, expert evaluation, or comparison against a non-interpretable baseline is provided. The paper therefore does not merely lack an evaluation of interpretability; it never states what the proposed interpretability method is.
  2. [Sections 4.2 and 4.3] The test-set construction invalidates the reported accuracy as an estimate of generalization. Section 4.2 states that 'some images were removed from the test and validation sets that are visually blurry and far-away but were still used in training,' and Section 4.3 reports that images with noise, poor contrast, nearby moles, and color lighting were manually removed, reducing the data from 10,015 to 7,115 images (about 28% discarded). Because the filtering rule is applied selectively to test and validation data, the reported test accuracy measures performance on a hand-picked subset rather than on the distribution of images encountered in a clinical or field deployment. The Bangladesh-specific deployment claim therefore has no supporting data.
  3. [Section 4.1 (Reflective noise removal) and Section 7] The preprocessing thresholds TR1 and TR2 in Eq. (1) are stated to be 'obtained by experimental' means, with no indication of a held-out tuning procedure or independent validation. Since these thresholds affect which images are retained and how artifacts are removed, tuning them on the same data used for evaluation introduces a form of information leakage that is not accounted for. Additionally, all reported performance numbers in Tables 2-3 and the AUC of 0.847 in Section 7 are point estimates without confidence intervals, error bars, or significance tests, so the claimed superiority of the proposed DCNN over the baselines is not statistically supported.
minor comments (5)
  1. [Section 4.2 (Data preparation) and Section 4.3 (Data reduction)] The sentence in Section 4.2 that mentions removing images from test and validation sets while keeping them in training is written as a statement of fact with a citation, but the methodological implications are not discussed. Please either remove this practice or justify it with a principled argument for why the filtered test set reflects the target deployment distribution.
  2. [Section 4 (Proposed methodology)] The subsection numbering is inconsistent: Sections 4.1, 4.2, and 4.3 appear multiple times, and the text after Section 4.3 reuses the same numbers (e.g., '4.1 Reflective noise and artifacts removal' after '4.3 Preprocessing'). This makes it difficult to refer to specific preprocessing steps. Renumber the subsections consecutively.
  3. [Equations in Section 6] Equations (6)-(9) for precision, recall, F1, and accuracy are correct, but there are no equations (3)-(5) after Eq. (2), and the references to 'Eq. (6)' in the surrounding text are not tied to the manuscript's equation numbering. Re-number the equations or add the missing intermediate equations.
  4. [Tables 2 and 3] The table columns are difficult to parse because the precision, recall, F1-score, training accuracy, and testing accuracy values are not clearly separated into rows per model. In Table 2, for example, the proposed DCNN row has six numeric values while the table has five columns of results. Please restructure the tables with one model per row and clear alignment of metrics.
  5. [Section 8 (Conclusion)] The conclusion states that 'there is no existing research that we are aware of that classified skin lesions as benign or malignant using the same dataset,' which is factually incorrect given the extensive literature on HAM10000 binary classification. Please remove or rephrase this claim and instead cite prior work on this dataset.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: reported accuracy is benchmarked on the external HAM10000 dataset, and the unsupported interpretability claim is an omission, not a circular reduction.

full rationale

The paper contains no derivation that reduces to its own inputs. The accuracy claim is an empirical comparison on the public HAM10000 dataset, and the transfer-learning baselines are external architectures trained on the same data. The preprocessing thresholds TR1 = 0.87 and TR2 = 0.096 are described as 'obtained by experimental' (Section 4.1), and the AUC of 0.847 is reported 'after multiple iterations of fine-tuning' (Section 7); these are ordinary model-selection choices, not fitted quantities renamed as predictions. The manual removal of blurry and far-away images from the test and validation sets while retaining them in training (Sections 4.2 and 4.3) undermines the validity of the reported generalization estimate, but it is a test-set curation problem, not a circular argument. The abstract and title promise a combination of saliency maps and attention maps, yet the body never specifies how these maps are computed, displays any map, or evaluates interpretability; this is a missing contribution and a falsifiability failure, not a self-referential derivation. No load-bearing self-citation appears: the cited work on data removal (Yan et al., 2023) and on interpretable skin-cancer classification (Mridha et al., 2023) is external to the authors. Therefore, by the strict definition used here, the paper has no significant circularity, and the appropriate score is 0.

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

The central metrics rest on the HAM10000 label set, a manually reduced image collection, and hand-tuned preprocessing thresholds. No external validation, formal derivation, or independent test population is provided.

free parameters (4)
  • Reflection threshold TR1 = 0.87
    Set experimentally in Eq. 1 on the same dataset; affects which pixels are classified as reflection artifacts.
  • Reflection threshold TR2 = 0.096
    Set experimentally in Eq. 1 on the same dataset; no sensitivity analysis is provided.
  • Data augmentation ranges = rotation 10, shifts 0.2, shear 0.2, zoom 0.2, channel shift 10
    Hand-picked values in Table 1 with no ablation study.
  • Manual data retention rule = 6,136 benign and 979 malignant images retained from 10,015
    Subjective deletion of blurry, artifact-heavy, or distant images; the rule is not specified enough to reproduce.
assumptions (4)
  • domain assumption HAM10000 ground truth labels are accurate for benign versus malignant classification.
    Used as the only supervision source in Section 4.1; any label noise transfers to the reported metrics.
  • ad hoc to paper Removing 28 percent of the images preserves the input distribution relevant for Bangladeshi deployment.
    No evidence that the retained images generalize to Bangladeshi skin tones or clinical settings; this is the load-bearing filter assumption.
  • domain assumption Pre-trained ImageNet weights transfer to dermoscopic images.
    The transfer learning baselines in Section 5 rely on this without domain adaptation checks.
  • standard math Sigmoid output with binary cross-entropy is appropriate for the two-class setup.
    A standard choice for binary classification, stated in Section 6 but not justified in context.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Using Computer Vision for Skin Disease Diagnosis in Bangladesh Enhancing Interpretability and Transparency in Deep Learning Models for Skin Cancer Classification." pith.science (2026). https://pith.science/paper/OTCICNF3

@misc{pith2026250118161,
  author       = {Pith},
  title        = {Pith review of: Using Computer Vision for Skin Disease Diagnosis in Bangladesh Enhancing Interpretability and Transparency in Deep Learning Models for Skin Cancer Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OTCICNF3}},
  note         = {Machine review of arXiv:2501.18161}
}
read the original abstract

With over 2 million new cases identified annually, skin cancer is the most prevalent type of cancer globally and the second most common in Bangladesh, following breast cancer. Early detection and treatment are crucial for enhancing patient outcomes; however, Bangladesh faces a shortage of dermatologists and qualified medical professionals capable of diagnosing and treating skin cancer. As a result, many cases are diagnosed only at advanced stages. Research indicates that deep learning algorithms can effectively classify skin cancer images. However, these models typically lack interpretability, making it challenging to understand their decision-making processes. This lack of clarity poses barriers to utilizing deep learning in improving skin cancer detection and treatment. In this article, we present a method aimed at enhancing the interpretability of deep learning models for skin cancer classification in Bangladesh. Our technique employs a combination of saliency maps and attention maps to visualize critical features influencing the model's diagnoses.

Figures

Figures reproduced from arXiv: 2501.18161 by the authors.

Figure 1
Figure 1. Challenges of skin lesions detection: a: hair artifacts, b: low contrast, c: irregular boundaries, d: color illumination.. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) 1. The pictures of skin cancer come in a variety of sizes and forms, making it difficult to identify the disease accurately. According to this viewpoint, pre-processi… view at source ↗
Figure 2
Figure 2. The following steps of our proposed methodology. 4.3. Preprocessing The first and most important stage in preparing the raw data and making it suitable with a machine learning model is data preparation. Preprocessing our dataset's primary goal is to improve the original medical pictures by eliminating air bubbles, noise, and artifacts that are brought on by gel that was applied before to image capture. To achieve a … view at source ↗
Figure 3
Figure 3. Removing noise and artifacts from the input image. 4.1. Reflective noise and artifacts removal To eliminate even a little amount of noise, there are several kinds of picture smoothing methods available, such as median blurring and Gaussian blurring. A basic thresholding method is employed to remove noise and artifact reflection from the photos. The elimination of noise from the original input picture is shown in [P… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Model architecture of our proposed DCNN model [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Regarding performance, confusion matrix of the proposed DCNN model [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: B1, B2, and B3 are benign, and M1, M2, and M3 are malignant skin lesions, respectively [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Performance curve based on accuracy per epoch of our proposed DCNN model [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Performance curve based on loss per epoch of our proposed DCNN model [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Receiver Operating Characteristic curve of our proposed DCNN model. 8 CONCLUSION AND FUTURE WORK In this study, the suggested DCNN model outperforms existing transfer learning models in terms of classification accuracy. The suggested technique has the capacity to categ…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

8 extracted references · 7 canonical work pages

  1. [1]

    According to this viewpoint, pre-processing is necessary for precise analysis

    The pictures of skin cancer come in a variety of sizes and forms, making it difficult to identify the disease accurately. According to this viewpoint, pre-processing is necessary for precise analysis

  2. [2]

    Section 4 details the materials and procedures

    The difficulties in detecting skin cancer are discussed in Section 3. Section 4 details the materials and procedures. In Section 5, we explain transfer learning using our suggested DCNN model. The training and performance of our suggested DCNN model are shown in Section 6. The results and discussion are presented in Section 7, and the conclusion and sugge...

  3. [3]

    Therefore, all of this noise and artifacting should be eliminated during the pre-processing processes

    In order to get a satisfactory outcome, a few inefficient signals that were not originally a component of the picture must be sacrificed. Therefore, all of this noise and artifacting should be eliminated during the pre-processing processes

  4. [4]

    Page 5 of 18

    Low contrast from nearby tissues might sometimes provide additional challenges and hinder accurate analysis of skin cancer. Page 5 of 18

  5. [5]

    Color lighting presents additional challenges because of its effects on color texture, light beams, and reflections

  6. [6]

    Although certain moles on the human body may never grow into cancer cells, they make it more difficult to effectively identify skin cancer from carcinogenic photos

  7. [7]

    4 PROPOSED METHODOLOGY In this part, we outline the subsequent phases and our methodology's flowchart, which is seen in Fig

    Another issue with detecting skin cancer is the present bias, which alters the performance of the m odels to provide a better result. 4 PROPOSED METHODOLOGY In this part, we outline the subsequent phases and our methodology's flowchart, which is seen in Fig. 2. When classifying benign from malignant lesions, DCNN and transfer learning models are utilized....

  8. [9]

    Additionally, the early identification of a malignant skin growth may greatly increase the chance of survival, particularly for people who lack access to medical care

    A CAD system that uses the suggested DCNN model may efficiently identify skin lesions at an early stage. Additionally, the early identification of a malignant skin growth may greatly increase the chance of survival, particularly for people who lack access to medical care. Table 2 The performance comparison between our proposed DCNN model and transfer lear...

Pith tools

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