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

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• Melanoma diagnosis relies on early detection, often hindered by limitations of traditional supervised learning methods requiring extensive labeled data.
• Supervised learning models for melanoma are sensitive to hyperparameter settings, impacting diagnostic accuracy.
• Conventional semi-supervised generative adversarial networks (SS-GANs) face challenges like mode collapse and poor generalization.

Purpose of the Study:

• To develop an enhanced diagnostic model for melanoma detection using a semi-supervised generative adversarial network (SS-GAN).
• To improve the accuracy and generalization of SS-GANs by addressing mode collapse, global dependency modeling, and pseudo-label reliability.
• To optimize model hyperparameters using an enhanced artificial bee colony (ABC) algorithm for robust melanoma classification.

Main Methods:

• Implemented a novel SS-GAN architecture incorporating reconstruction loss and self-attention mechanisms in both generator and discriminator.
• Applied consistency regularization and confident pseudo-labeling strategies to enhance discriminator stability and training with unlabeled data.
• Utilized a mutual learning-based artificial bee colony (ML-ABC) algorithm for hyperparameter optimization, reducing reliance on manual tuning.

Main Results:

• The proposed model achieved high F-measures across four benchmark datasets: ISIC-2020 (92.77%), HAM10000 (93.38%), PH2 (90.63%), and DermNet (92.62%).
• Demonstrated significant improvement in distinguishing melanoma from non-melanoma images, particularly under conditions of limited labeled data.
• The enhanced SS-GAN effectively addressed mode collapse and improved feature representation through self-attention and reconstruction loss.

Conclusions:

• The developed ML-ABC optimized SS-GAN provides a robust and accurate approach for melanoma image classification, especially with scarce labeled data.
• The proposed architectural improvements and optimization strategy enhance the reliability and performance of generative adversarial networks in medical diagnostics.
• This study offers a valuable tool for early melanoma detection, contributing to improved patient outcomes and treatment efficacy.