Adapting a style based generative adversarial network to create images depicting cleft lip deformity.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Scarcity of high-quality, ethics-approved patient images impedes machine learning for facial deformity evaluation.
• Developing robust machine learning models requires extensive and diverse datasets.

Purpose of the Study:

• To develop a deep learning-based system (CleftGAN) for generating a large, varied dataset of realistic cleft lip facial images.
• To assess the performance of CleftGAN in generating high-fidelity facial image facsimiles.

Main Methods:

• Utilized a transfer learning protocol with StyleGAN as the base model.
• Employed data augmentation with 514 cleft lip images and a base model of 70,000 normal faces.
• Evaluated generated images using Frechet Inception Distance, Perceptual Path Length, and Divergence Index of Normality.

Main Results:

• CleftGAN successfully generated a vast number of unique, high-fidelity cleft lip images with ethnic diversity.
• Performance metrics confirmed high similarity to the training dataset and semantically valid image interpolation.
• Generated images exhibited a comparable distribution of normality to the training data.

Conclusions:

• CleftGAN is a novel instrument for generating an almost unlimited supply of realistic cleft lip facial images.
• This tool can significantly aid the development of machine learning models for objective facial form evaluation and surgical outcome assessment.