A hybrid framework for Turner Syndrome facial detection: integrating deep feature learning and geometric prior knowledge under data scarcity

Abstract Turner Syndrome (TS) is a chromosomal disorder characterized by growth retardation and congenital ovarian dysgenesis. Given the considerable morbidity and mortality, early diagnosis and timely intervention are crucial to improving outcomes. However, the gold standard diagnostic method, karyotype analysis, is time-consuming and requires specialized laboratory expertise. Since TS patients often present distinct facial features, AI-based facial recognition could help identify patients with characteristic appearance as a screening tool. Inspired by the value-of-information principle, we propose a hybrid framework integrating deep feature learning and geometric prior knowledge under data scarcity. We retrospectively constructed this AI facial recognition model for TS detection and validated it in a prospective cohort. The proposed framework achieved 95.41% accuracy, 96.55% sensitivity, and 95.24% specificity, substantially simplifying the diagnostic process. When training was performed only on the hard subset obtained after difficulty-based stratification into three equal groups, the sensitivity reached 100.00%, indicating that all patients were correctly identified. Grad-CAM visualization of key facial regions improved model interpretability and aligned predictions with clinical insights. Our method offers a low-cost and easily accessible approach to TS screening. By allowing for an earlier detection of TS, particularly in medical resource-limited settings, this method has the potential to facilitate timely interventions. This novel paradigm also provides insight for facial recognition-aided diagnosis in other diseases.