Semantic Fidelity Over Speed: A Comparative Study of Conditional vs. Unconditional GANs for Few-Shot Skin Disease Classification

Abstract

This research aims to break the crucial bottleneck of data deficiency in developing AI for skin disease diagnosis by presenting a clear and evidence-based comparison of two different generative AI models for data augmentation. By using a small clinical data set, this research systematically compared a fast but unconditional model of generative AI, FastGAN, with a semantic-aware model of generative AI, cGAN. Synthetic data generated by these models was used for training and testing different classifiers for different kinds of medical diagnoses. The results of this research clearly show that there is a crucial dependency of AI models for medical data augmentation on semantic fidelity. The cGAN model, which is semantic-aware and preserves class-specific features of skin diseases, enabled classifiers to retain high accuracy (up to 93%) for different kinds of medical diagnoses. However, in stark contrast, the unconditional FastGAN model, despite being much faster in generating synthetic data, catastrophically failed in retaining accuracy as low as 49% due to semantic inconsistency in data augmentation.