Identifying suspicious naevi with dermoscopy via variational autoencoder auxiliary generative classifiers

Area of Science:

• Dermatology and Artificial Intelligence
• Computational Biology and Bioinformatics

Background:

• Naevi (moles) are benign skin tumors requiring study due to their link with melanoma risk.
• Accurate classification of naevi is crucial for early melanoma detection and patient outcomes.

Purpose of the Study:

• To create a visual manifold illustrating the distribution of suspicious and non-suspicious naevi.
• To classify real naevi and generate realistic synthetic samples using generative adversarial networks (GANs).
• To apply data-driven methods for early melanoma detection by identifying distinct features of suspicious naevi.

Main Methods:

• Utilized a variational autoencoder auxiliary classifier generative adversarial network (VAE-ACGAN) for naevi classification.
• Generated realistic synthetic naevi images and interpreted their distribution via a variational manifold.
• Compared the VAE-ACGAN model's performance against various deep learning frameworks.

Main Results:

• The VAE-ACGAN model achieved outstanding performance in specificity, sensitivity, and AUC scores, particularly for suspicious naevi.
• The generated manifold clearly distinguished between suspicious and non-suspicious naevi categories.
• The models produced high-quality, life-like representations of naevi, outperforming other deep learning frameworks.

Conclusions:

• GANs show significant potential for expanding dermatological datasets and improving deep learning algorithm effectiveness.
• Interpretable clustering based on visual similarities can enhance naevi categorization and understanding.
• Accurate naevi identification and classification using AI can facilitate earlier melanoma detection.