Boosting Skin Lesion Classification with a Class Expert DCGAN Framework for Skin Disease Detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models for skin lesion classification face challenges with limited data and imbalanced datasets.
• Underrepresented classes in datasets hinder accurate classification and diagnostic reliability.

Purpose of the Study:

• To introduce a novel framework, the class expert Deep Convolutional Generative Adversarial Network (DCGAN), to address class imbalance and improve classification accuracy for underrepresented skin lesion classes.
• To leverage weight transfer from a GAN discriminator to enhance Convolutional Neural Network (CNN) performance through discriminative feature extraction.

Main Methods:

• Developed a class expert DCGAN framework incorporating weight transfer from a GAN discriminator to expert layers.
• Utilized transfer learning by applying discriminator weights to CNN model layers for improved feature learning.
• Conducted experimental evaluations to assess the framework's performance on skin lesion classification tasks.

Main Results:

• The class expert DCGAN framework demonstrated notable improvements in accuracy and precision, especially for classes with fewer samples.
• Achieved a 2-3% increase in classification accuracy compared to traditional skin lesion classification methods.
• Validated the effectiveness of GANs for data augmentation and discriminative feature extraction in medical image analysis.

Conclusions:

• The proposed class expert DCGAN framework provides an effective solution for improving skin lesion classification performance.
• This approach enhances diagnostic reliability and facilitates better interpretation of dermatological images in clinical settings.
• Leveraging GANs for data augmentation and feature extraction shows significant potential in medical image classification challenges.