ETD

• data augmentation
• generative AI
• image classification
• oscc
• stable diffusion
• styleGAN3

This thesis explores how humans perceive synthetic medical images generated for oral cancer applications. Specifically, it evaluates the realism and usefulness of synthetic photographic images of oral squamous cell carcinoma (OSCC) lesions, previously produced using two state-of-the-art generative frameworks — StyleGAN3 and Stable Diffusion — under different conditioning strategies. The study draws on two OSCC photographic datasets, including the publicly available PhotoMOCI dataset. Human evaluation involved both medical experts and non-expert participants, who were asked to distinguish real images from synthetic ones and to judge how realistic each image appeared. The analysis then examines how this human perception relates to the effectiveness of synthetic images when used for data augmentation in lesion classification tasks, performed with deep learning models such as ResNet50 and the Vision Transformer. The findings reveal that different generation techniques lead to markedly different levels of perceived realism, and that images judged more realistic by human observers do not necessarily translate into better classification performance.