An Enhanced Approach Using AGS Network for Skin Cancer Classification.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer is a prevalent global health issue, with diagnosis often hindered by visual similarities between subtypes.
• Dermatologists' diagnostic accuracy for skin cancer ranges from 62% to 80%, highlighting a need for improved tools.
• Training AI models for skin cancer classification requires extensive medical image datasets, which are challenging to acquire.

Purpose of the Study:

• To propose the AGS network, an integrated framework combining Augmentation, GAN, and Segmentation modules.
• To enhance the performance of skin cancer classification using AI, particularly in scenarios with limited data.
• To evaluate the effectiveness of the AGS network across multiple deep learning classifiers.

Main Methods:

• The AGS network was developed, integrating Augmentation (A), Generative Adversarial Network (G), and Segmentation (S) modules.
• Eight deep learning classifiers (GoogLeNet, DenseNet201, ResNet50, MobileNet V3, EfficientNet B0, ViT, EfficientNet V2, Swin Transformers) were evaluated.
• The HAM10000 dataset was used for evaluation, with five model configurations tested to assess module contributions.

Main Results:

• All eight classifiers showed consistent performance improvements when integrated with the AGS network.
• EfficientNet V2 + AGS demonstrated the largest gains, with +0.1808 in Accuracy and +0.1674 in F1-Score over baseline models.
• ResNet50+AGS achieved the highest overall performance, reaching 95.87% Accuracy and 95.73% F1-Score.

Conclusions:

• The AGS network effectively addresses the challenge of small datasets in skin cancer classification.
• Combining multiple augmentation techniques within an integrated framework significantly boosts AI classifier performance.
• The study validates the AGS network as a powerful tool for improving skin cancer diagnostic accuracy.