AI-Based Data Augmentation for Cleft Lip Assessment: Generating Realistic Synthetic Videos to Improve Clinical Outcomes

Cleft lip and/or palate (CLP) represents one of the most prevalent congenital anomalies, exerting significant functional and aesthetic impact during childhood. Despite the growing interest in artificial intelligence (AI) applications for medical diagnostics, the development of AI-based tools for CLP assessment is hindered by the limited availability of clinical data, mainly due to ethical and logistical constraints. This study investigates the potential of generative models to synthesize realistic video sequences of children with CLP from static images, aiming to augment the pool of clinically relevant data. Two state-of-the-art AI-driven facial animation models were compared: the First Order Motion Model (FOMM) and LivePortrait (LP). Their performance was assessed using a combination of perceptual and geometric metrics, including the Learned Perceptual Image Patch Similarity (LPIPS, both global and lip-localized), Average Expression Distance (AED), and Normalized Mean Error (NME), under both self-animation and cross-animation scenarios. Additionally, the influence of clinical severity and phonetic complexity of spoken phrases on model performance was analyzed. Results indicate that LP consistently outperforms FOMM, generating animations that are more realistic, structurally coherent, and accurate in the nasolabial region, even in severe clinical cases. These findings support the use of LP as a promising tool for the synthetic generation of medical data, with potential applications in AI training, clinical education, and the development of automated evaluation systems.