What type of study is this?

September 10, 2025Open Access

Semi-supervised GAN with hybrid regularization and evolutionary hyperparameter tuning for accurate melanoma detection

Key Points

The model achieved high F-measure scores across multiple datasets, showcasing its robust detection capabilities.
Enhancements such as reconstruction loss and self-attention were critical in improving global feature modeling.
Adopting mutual learning-based optimization for hyperparameter tuning reduced sensitivity to parameter choices.
Validation on diverse melanoma datasets confirms the model's effectiveness in distinguishing melanoma from non-melanoma images.

Abstract

Melanoma, influenced by changes in deoxyribonucleic acid (DNA), requires early detection for effective treatment. Traditional melanoma research often employs supervised learning methods, which necessitate large, labeled datasets and are sensitive to hyperparameter settings. This paper presents a diagnostic model for melanoma, utilizing a semi-supervised generative adversarial network (SS-GAN) to enhance the accuracy of the classifier. The model is further optimized through an enhanced artificial bee colony (ABC) algorithm for hyperparameter tuning. Conventional SS-GANs face challenges such as mode collapse, weak modeling of global dependencies, poor generalization to unlabeled data, and unreliable pseudo-labels. To address these issues, we propose four improvements. First, we add a reconstruction loss in the generator to minimize mode collapse and maintain structural integrity. Second, we introduce self-attention in both the generator and the discriminator to model long-range dependencies and enrich features. Third, we apply consistency regularization on the discriminator to stabilize predictions on augmented samples. Fourth, we use pseudo-labeling that leverages only confident predictions on unlabeled data for supervised training in the discriminator. To reduce dependence on hyperparameter choices, the Random Key method is applied, enhanced through a mutual learning-based ABC (ML-ABC) optimization. We evaluated the model on four datasets: International Skin Imaging Collaboration 2020 (ISIC-2020), Human Against Machine's 10,000 images (HAM10000), Pedro Hispano Hospital (PH2), and DermNet datasets. The model demonstrated a strong ability to distinguish between melanoma and non-melanoma images, achieving F-measures of 92.769%, 93.376%, 90.629%, and 92.617%, respectively. This approach enhances melanoma image classification under limited labeled data, as validated on multiple benchmark datasets. Code is publicly available at https://github.com/AmirhoseinDolatabadi/Melanoma .

Read Full Paperexternally

Mark Helpful

Bookmark

Relay

View Full Paper