Skin Lesion Synthesis and Classification Using an Improved DCGAN Classifier

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Early detection of skin cancer significantly improves patient prognosis.
• Automated diagnostic technologies are crucial for early skin lesion detection.
• Challenges in skin lesion classification include limited annotated data and class imbalance, hindering deep learning model training.

Purpose of the Study:

• To propose a novel skin lesion synthesis and classification model using an Improved Deep Convolutional Generative Adversarial Network (DCGAN).
• To address the data scarcity and class imbalance issues in skin cancer datasets.
• To generate high-quality synthetic skin lesion images for improved automated diagnosis.

Main Methods:

• Development of an Improved DCGAN model for generating realistic synthetic skin lesion images.
• Application of image enhancement techniques including scaling, normalization, sharpening, color transformation, and median filters.
• Utilizing the Discriminator's final layer as a classifier for binary classification (benign vs. malignant).
• Training with a constant learning rate of 0.01 and optimized hyperparameters.

Main Results:

• The DCGAN Classifier model achieved superior performance on the ISIC2017 dataset.
• Achieved an accuracy of 99.38%, with 99% for recall, precision, F1 score, and Balanced Accuracy Score (BAS).
• Outperformed existing state-of-the-art deep learning models in skin lesion classification.

Conclusions:

• The proposed DCGAN Classifier effectively generates high-quality synthetic skin lesion images.
• The model demonstrates high accuracy in classifying benign and malignant skin lesions.
• This approach shows significant promise for deep learning-based medical image analysis in dermatology.