SynthVision - Harnessing Minimal Input for Maximal Output in Computer Vision Models using Synthetic Image Data

This paper introduces SynthVision, a novel approach to rapidly develop disease detection computer vision models using minimal real-world data. The approach leverages diffusion models to generate a large synthetic dataset of images, specifically focusing on detecting Human Papilloma Virus (HPV) genital warts. A two-phase experimental design was employed: Phase 1 fine-tuned diffusion models using 10 guide images to generate 500 clinically validated synthetic images; Phase 2 trained and tested a vision transformer model on this synthetic dataset. The model achieved exceptional performance, with 96% accuracy, 99% precision for HPV cases, 94% recall for HPV cases, 95% precision for normal cases, and 99% recall for normal cases. The F1-score was 96% for HPV and 97% for normal cases. SynthVision demonstrates the potential to develop accurate computer vision models from minimal input, offering a vital solution for rapid response to medical emergencies.

Yudara Kularathne, Prathapa Janitha, Sithira Ambepitiya, Thanveer Ahamed, Dinuka Wijesundara, Prarththanan Sothyrajah