Accelerating Facial Anomaly Appraisal: A Knowledge Distillation Approach

This study introduces a novel machine learning framework that aims to reliably detect, locate, and evaluate facial anomalies such as cleft lip deformities. The objective is to create a universal and objective method for assessing facial abnormalities, with sensitivity to both minor and significant deformities. The framework utilizes an efficient knowledge distillation model to generate an anomaly heatmap that identifies potential facial irregularities. This heatmap is then transformed into a score representing the severity of the facial deformity. The proposed approach achieves results on par with state-of-the-art methods but significantly outperforms them in terms of speed, requiring only 100 milliseconds from image upload to generating the facial rating. This efficiency makes it ideal for integration into mobile applications. Moreover, the method does not rely on anomalous data for training, yet it effectively detects and evaluates various facial anomalies. We demonstrate that this novel computerized system produces facial normality/abnormality scores that align closely with human judgment, achieving a 88% correlation with human ratings.Clinical relevance- Precise detection and measurement of facial anomalies by healthcare professionals and patients play a crucial role in clinical practice. These evaluations support pre-surgical planning, provide evidence to secure approval for surgical procedures from insurance companies, and facilitate unbiased assessments of post-surgical results, determining whether the surgery was successful or if additional corrections are necessary.