Improving Skin Cancer Classification Using Heavy-Tailed Student T-Distribution in Generative Adversarial Networks...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models require large datasets, which are often scarce in medical imaging.
• Limited medical data hinders the performance of deep learning algorithms in diagnostic tasks.

Purpose of the Study:

• To propose a novel framework (TED-GAN) for generating realistic medical images, specifically skin lesions.
• To address the bottleneck of insufficient data for deep learning in medical imaging.
• To improve the classification performance of skin lesion diagnosis.

Main Methods:

• A framework combining a variational autoencoder (VAE) and two generative adversarial networks (GANs) with an auxiliary classifier.
• Utilizing an encoder-decoder network to extract informative latent noise vectors.
• Employing a heavy-tailed student t-distribution for noise sampling in one GAN to enhance image diversity.
• Training the system to generate synthetic skin lesion images.

Main Results:

• The proposed TED-GAN framework successfully generated realistic-looking skin lesion images.
• Skin lesion classification accuracy improved from 66% to 92.5% using the generated images.
• The use of a t-distribution in the GAN contributed to diverse image generation and better classification impact.

Conclusions:

• TED-GAN effectively overcomes medical data scarcity for deep learning applications.
• The framework demonstrates significant potential for improving diagnostic accuracy in medical imaging tasks.
• The method offers a viable solution for augmenting datasets and enhancing model performance in areas with limited data.