Realistic generation of synthetic pediatric liver ultrasound images for liver disease classification

Oral presentation: Fake it till you make it: Generative AI models for creating realistic artificial pediatric liver ultrasound images.

Poster presentation: Fake It Till You Make It: Synthetic Generation of Pediatric Liver Ultrasound Images Using Generative Ai Models

Large-scale labeled biomedical datasets are difficult to access, and these datasets remain small in the medical imaging domain compared to other common computer visiom datasets. In this project, we prototype a method of using generative adversarial networks to generate synthetic liver ultrasound images. The preliminary results show that the synthetic images have good quality compared with the real images, and this synthetic dataset would be promising as a data augmentation method for liver disease classification task.

My contributions to this project include manual cleaning of approximately 1000 ultrasound scans, building codebase for ultrasound deep convolutional GAN (DCGAN), training StyleGAN, and evaluating model performance using moment methods and other metrics, such as Frechet Inception Distance (FID).

I’d like to thank Dr. Surya Prasath for the opportunity to work on this project!