PaperlayerView exact revision on arXiv

Source-linked AI summary

A Lightweight Attention-Based CNN with Saliency-Map Interpretability for Skin Cancer Detection

Sanjana Islam Kasfia, Md. Ismiel Hossen Abir, AKM Bahalul Haque, Abir Mahmud Shahariar

arXiv:2610.37314v1cs.CVcs.AI

TL;DR

Skin lesion diagnosis is challenging because lesions can look similar and automated classification must balance accuracy with computational demands. The paper combines a lightweight attention-based CNN with saliency maps for binary classification. It reports 87% accuracy, matching MobileNet while using fewer parameters and less storage.

  • Problem

    Skin lesions can look visually similar, vary within categories, and occur in imbalanced datasets, complicating reliable automated classification.

    Source: §I · pp. 1–2

  • Method

    The study uses a lightweight CNN-Attention model for binary classification and saliency maps to visualize image regions influencing predictions.

    Source: §I · p. 2, §VI · pp. 5–6

  • Results

    87% accuracy matched MobileNet; CNN-Attention used 1.95 million parameters and 7.45 MB storage, with ROC AUC 0.9475 and average PR-AUC 0.9430.

    Source: §IV · p. 4, §V · p. 5, §VI · p. 6

  • Takeaways & Limitations

    The study combines classification performance, lower model complexity, and prediction interpretability in one framework.

    Source: §VI · p. 6

Abstract

from arXiv · show

Skin cancer is among the most prevalent and life threatening forms of cancer worldwide. Early and accurate diagnosis significantly improves patient survival rates. However, manual dermoscopic assessment remains challenging due to visual similarity. This study proposes an Explainable AI-driven framework for automated binary classification of skin lesions using a lightweight attention-based convolutional neural network (CNN-Attention). The model integrates channel attention to focus on informative regions of dermoscopic images while maintaining low computational complexity. A publicly available dataset comprising 2,637 training and 660 test images was used, with balanced class weights. The proposed model achieved 87 percent accuracy with 1.95 million parameters and 7.45 MB of storage, matching the best baseline (MobileNet) while using approximately half as many parameters and half as much storage. ROC AUC and PR AUC scores reached 0.9475 and 0.9430, respectively. Additionally, saliency maps were incorporated as an explainability technique to visualize regions that influence model decisions. Comparative analysis with VGG16,VGG19,ResNet101,and MobileNet demonstrates the proposed models effectiveness in balancing classification results, computational efficiency, and interpretability.

I. INTRODUCTION

Skin lesion diagnosis is difficult because lesions can look alike, vary widely, and appear in images with artifacts, while conventional assessment can be time-consuming and expert-dependent. The study proposes a lightweight attention-based CNN and saliency maps for binary lesion classification.

I. INTRODUCTION

  • Skin cancer includes malignant tumors arising from abnormal skin-cell growth, and melanoma is described as its most dangerous form.Early identification of malignant lesions is presented as critical to successful treatment and improved survival and recovery prospects.Source: §I · p. 1
  • Lightweight CNNs must balance classification accuracy, attention to salient features, and reduced computational demands.The introduction identifies class imbalance, visual similarity, within-class variation, and image artifacts as challenges for automated classification.Source: §I · p. 2
  • The framework combines attention-based CNN classification with saliency maps that visualize regions influencing model decisions.The study also compares its proposed model with pretrained architectures using standard evaluation metrics.Source: §I · p. 2

II. LITERATURE REVIEW

The literature review summarizes current approaches to skin lesion identification, including feature-extraction techniques and learning architectures.

II. LITERATURE REVIEW

  • The review surveys current research approaches to skin lesion identification.The supplied passage identifies the section as a summary of current research approaches.Source: §II · p. 2
  • Feature extraction techniques are among the approaches covered in the review.Source: §II · p. 2
  • The review also covers learning architectures used for skin lesion identification.Source: §II · p. 2

A. Traditional and Feature-Based Methods

The review spans traditional feature-based systems and deep-learning approaches, including transfer learning and hybrid models.

A. Traditional and Feature-Based Methods

  • Salomi et al. combined CNNs, ABCD rule-based features, and K-means clustering for lesion classification.Their system also included an Android application; Hatem et al. used MATLAB and K-nearest neighbors to classify lesions as normal or malignant.Source: A. Traditional and Feature-Based Methods · p. 2

B. Deep Learning and Transfer Learning

  • On HAM10000, Shetty et al.'s custom CNN achieved 95.18% accuracy, while Bozkurt et al. reported improvement from 83.59% to 95.09% with an augmented dataset.Source: B. Deep Learning and Transfer Learning · p. 2
  • Other reviewed approaches include a patch-processing LSTM-CNN for binary classification, an ANN/CNN review, and a framework combining image fusion, segmentation, ResNet-50, and feature optimization,,.Source: B. Deep Learning and Transfer Learning · p. 3

C. Advanced and Efficient Deep Learning Approaches

The review describes efficient deep-learning approaches before outlining the study workflow and dataset preparation for binary lesion classification.

C. Advanced and Efficient Deep Learning Approaches

  • Reviewed efficient approaches include SKINC-NET for seven lesion types, segmentation with adaptive CNNs, and hybrid or optimized CNN designs,,,,.The review also mentions quantum machine-learning approaches evaluated against conventional pretrained models.Source: C. Advanced and Efficient Deep… · p. 3

III. METHODOLOGY

  • The methodology describes the overall CNN-with-attention workflow for skin cancer classification.Figure 1 is identified as the study workflow diagram.Source: §III · p. 3

A. Data Analysis and Preparation

  • The public binary-classification dataset contained 2,637 training images (1,440 benign and 1,197 malignant) and 660 test images; training imbalance was addressed using class weights.The training data was split stratified, with 85% for training and 15% for validation; class weights were approximately 0.915 for benign and 1.102 for malignant.Source: A. Data Analysis and Preparation · p. 3

B. Proposed CNN-Attention Model

The proposed lightweight CNN-Attention model combines four convolutional blocks with channel attention and fully connected classification layers. Its architecture is presented in Figure 3.

B. Proposed CNN-Attention Model

C. Transfer Learning

Four pretrained CNNs were adapted for binary skin-lesion classification using ImageNet weights and a shared classifier-head design.

C. Transfer Learning

  • VGG16, VGG19, ResNet101, and MobileNet were adapted as pretrained baselines for binary classification.Source: C. Transfer Learning · p. 4
  • Each model used an ImageNet-weighted backbone followed by global average pooling and two dense layers of 512 and 256 units.Source: C. Transfer Learning · p. 4
  • The classifier heads included batch normalization, dropout of 0.5 and 0.3, and a softmax output layer.Source: C. Transfer Learning · p. 4

D. Explainable AI and Evaluation Metrics

The evaluation combines saliency-map explanations with classification and discrimination metrics, while training uses early stopping and learning-rate scheduling.

D. Explainable AI and Evaluation Metrics

IV. EXPERIMENTAL RESULTS

The CNN-Attention model achieved 0.87 accuracy, with strong ROC and PR AUC scores across the reported classes. Saliency visualizations show model focus on lesion regions.

IV. EXPERIMENTAL RESULTS

  • 0.87 accuracy was reported, with benign precision, recall, and F1-score of 0.91, 0.85, and 0.88 and malignant values of 0.83, 0.89, and 0.86.Macro and weighted averages were both 0.87.Source: §IV · p. 4, Table I · p. 4
  • 307 benign and 268 malignant cases were correctly predicted, alongside 53 false positives and 32 false negatives.Source: §IV · p. 4, Figure 4 · p. 5
  • 0.9475 ROC AUC was reported for both benign and malignant classes.Source: §IV · p. 4, Figure 5 · p. 5
  • 0.9430 average PR-AUC was reported, with 0.9563 for benign and 0.9296 for malignant cases.Source: §IV · p. 4, Table II · p. 5
  • Saliency maps for benign and malignant examples show bright regions focused on the lesion rather than surrounding skin.The reported confidence scores were 98.9% for the benign example and 71.3% for the malignant example.Source: §IV · p. 5, Figure 6 · p. 5, Figure 7 · p. 5

V. DISCUSSION AND COMPARATIVE ANALYSIS

CNN-Attention matched MobileNet’s 87% accuracy while using fewer parameters and less storage. It also outperformed VGG16, VGG19, and ResNet101 on accuracy.

V. DISCUSSION AND COMPARATIVE ANALYSIS

  • 84% accuracy was achieved by both VGG16 and VGG19, while ResNet101 reached 83% and MobileNet achieved 87%.Source: §V · p. 5
  • 43.84 million parameters made ResNet101 the largest model; CNN-Attention used 1.95 million parameters and 7.45 MB, compared with MobileNet’s 3.89 million parameters and 14.84 MB.Source: §V · p. 5
  • 87% accuracy matched MobileNet, with CNN-Attention using 1.95 million parameters and 7.45 MB of storage.Table III compares model performance and complexity.Source: Table III · p. 5, §V · p. 5

VI. CONCLUSION

The conclusion presents a lightweight CNN-Attention model for benign-versus-malignant classification, pairing lesion-focused attention with saliency-map explanations. It reports 87% accuracy, matching MobileNet while using fewer parameters and less storage.

VI. CONCLUSION

  • 87% accuracy matched MobileNet, with 1.95 million parameters, 7.45 MB of storage, and an average PR-AUC of 0.9430.Precision was 0.91 for benign and 0.83 for malignant.Source: §VI · p. 6
  • The attention mechanism focuses on important skin-lesion regions, while saliency maps show image regions that influenced classification decisions.Source: §VI · pp. 5–6
  • The framework combines classification performance, lower model complexity, and prediction interpretability.Source: §VI · p. 6

Related papers

Loading 2610.37314v1…