Multi-Class Skin Problem Classification Using Deep Generative Adversarial Network (DGAN)

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Dermatology

Background:

• Automatic detection of skin problems is hindered by a lack of annotated datasets.
• Deep learning models require large volumes of labeled data, which are often unavailable.
• Traditional data augmentation methods have limitations in addressing data scarcity.

Purpose of the Study:

• To develop a deep generative adversarial network (DGAN) for multi-class classification of skin problems.
• To generate synthetic skin images to augment limited datasets.
• To improve the stability and performance of DGAN models in medical applications.

Main Methods:

• Developed a multi-class DGAN classifier to learn data distribution and generate synthetic skin images.
• Integrated data from diverse online sources to address class imbalance.
• Trained and evaluated Convolutional Neural Network (CNN) models (ResNet50, VGG16) using traditional augmentation for comparison.

Main Results:

• The DGAN model achieved high performance: 91.1% on unlabelled datasets and 92.3% on labelled datasets.
• CNN models with traditional data augmentation achieved a maximum performance of 70.8% on unlabelled datasets.
• DGAN demonstrated superior performance compared to conventional data augmentation techniques.

Conclusions:

• DGAN is a viable and effective solution for generating synthetic medical images, addressing data scarcity in dermatology.
• The developed DGAN model can learn from unlabelled datasets to produce accurate diagnostic results.
• This approach significantly enhances the diagnostic accuracy for skin problems, even with limited annotated data.