Home / Current Issue / Paper 1717366
Confidence-Aware Dual-CNN Framework for Vitiligo Detection Using EfficientNet-B0 with Freeze-Unfreeze Transfer Learning
Subject area: Science,Engineering and Technology · Area of research: Artificial Intelligence And Machine Learning
DOI: https://doi.org/10.64388/IREV9I11-1717366
Abstract
Vitiligo is a chronic depigmentation disorder affecting approximately 1–2% of the global population, and its accurate automated detection remains a challenging task due to variations in lesion size, skin tone, and imaging conditions. Existing deep learning-based vitiligo classifiers produce binary predictions without any measure of confidence or uncertainty, which can be clinically misleading. This paper proposes a Confidence-Aware Dual-CNN framework that employs two EfficientNet-B0 models with architecturally distinct classification heads, trained using a freeze-unfreeze transfer learning strategy. When both models produce consistent predictions, the system outputs a confident diagnosis of either Healthy or Vitiligo. When the models disagree beyond a designed disagreement threshold, the case is flagged as Uncertain and referred to a dermatologist for manual review. This is the first vitiligo detection framework to incorporate disagreement-based clinical uncertainty estimation. Evaluated on a dataset of 3,628 images, Model A achieves 94.98% accuracy, Model B achieves 96.28% accuracy, and the Dual-CNN framework achieves 97.82% accuracy on confident predictions, surpassing the existing IEEE CNN Autoencoder baseline of 90.16%. Supporting modules include an OpenCV-based lesion segmentation pipeline for progression tracking and a Random Forest model for treatment recommendation achieving 90% accuracy with a macro F1-score of 0.87.
Keywords
Vitiligo Detection, Dual CNN, EfficientNet-B0, Transfer Learning, Uncertainty Estimation, Disagreement Threshold, OpenCV, Random Forest, Medical Image Classification
References
[1] S. A. Alzakari et al., "LesionNet: An automated approach for skin lesion classification using SIFT features with customized CNN," Frontiers in Medicine, vol. 11, 2024.
[2] P. N. Srinivasu et al., "Classification of Skin Disease Using Deep Learning Neural Networks with MobileNet V2 and LSTM," Sensors, vol. 21, no. 8, p. 2852, 2021.
[3] J. Peng et al., "Classification of Non-tumorous Facial Pigmentation Disorders Using GAN and Improved SMOTE," in Proc. 43rd IEEE EMBC, 2021, pp. 1–4.
[4] "Vitiligo Binary Classification Using CNN Autoencoder," IEEE Xplore, 2022. [Accuracy: 90.16%]
[5] R. R. Selvaraju et al., "Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization," in Proc. IEEE ICCV, 2017, pp. 618–626.
[6] G. Munjal et al., "SkinSage XAI: An explainable deep learning solution for skin lesion diagnosis," Health Care Science, vol. 3, pp. 438–455, 2024.
[7] M. Tan and Q. V. Le, "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks," in Proc. ICML, 2019.
[8] H. Liu et al., "A skin disease classification model based on multi-scale combined efficient channel attention module," Scientific Reports, vol. 15, 2025.
How to cite this paper
@article{1717366,
author = {Dheeraj S Kumar, Rohit John Alex, Adhidev M D, Dr. K. Arthi},
title = {Confidence-Aware Dual-CNN Framework for Vitiligo Detection Using EfficientNet-B0 with Freeze-Unfreeze Transfer Learning},
journal = {Iconic Research And Engineering Journals},
year = {2026},
volume = {9},
number = {11},
pages = {1406-1412},
issn = {2456-8880},
url = {https://www.irejournals.com/formatedpaper/1717366.pdf},
abstract = {Vitiligo is a chronic depigmentation disorder affecting approximately 1–2% of the global population, and its accurate automated detection remains a challenging task due to variations in lesion size, skin tone, and imaging conditions. Existing deep learning-based vitiligo classifiers produce binary predictions without any measure of confidence or uncertainty, which can be clinically misleading. This paper proposes a Confidence-Aware Dual-CNN framework that employs two EfficientNet-B0 models with architecturally distinct classification heads, trained using a freeze-unfreeze transfer learning strategy. When both models produce consistent predictions, the system outputs a confident diagnosis of either Healthy or Vitiligo. When the models disagree beyond a designed disagreement threshold, the case is flagged as Uncertain and referred to a dermatologist for manual review. This is the first vitiligo detection framework to incorporate disagreement-based clinical uncertainty estimation. Evaluated on a dataset of 3,628 images, Model A achieves 94.98% accuracy, Model B achieves 96.28% accuracy, and the Dual-CNN framework achieves 97.82% accuracy on confident predictions, surpassing the existing IEEE CNN Autoencoder baseline of 90.16%. Supporting modules include an OpenCV-based lesion segmentation pipeline for progression tracking and a Random Forest model for treatment recommendation achieving 90% accuracy with a macro F1-score of 0.87.},
keywords = {Vitiligo Detection, Dual CNN, EfficientNet-B0, Transfer Learning, Uncertainty Estimation, Disagreement Threshold, OpenCV, Random Forest, Medical Image Classification},
month = {May},
doi = {https://doi.org/10.64388/IREV9I11-1717366}
}