Deep learning model for the development of visual parameters for upper lip assessing in patients operated on for unilateral cleft lip

Rossell-Perry P.; Arias-Figueroa J. · Cirugia Plastica Ibero-Latinoamericana (2024) · DOI: 10.4321/s0376-78922024000400012

Resumen

Background and objective. Cleft lip and palate is a congenital defect (the second most frecuente congenital malformation) of the face that affects the formation of the upper lip, alveolar and palatine region. The social impact that this pathology represents is great; patients are affected psychologically and functionally, and may present speech and eating defectsas well as aesthetic consequences that generate anxiety, depression, low self-esteem and others. The main objective of our study is to develop a deep learning model that allows assessing postoperative upper labial visual parameters in patients operated on for unilateral cleft lip. Methods. Postoperative images of 364 patients operated on for unilateral cleft lip were collected and rated by 3 expert reviewers as good or poor results. Three convolutional neural network models were trained to determine lip symmetry and postoperative outcome quality and were compared based on their sensitivity, specificity, and ROC-AUC curve. Results. The 3 convolutional neural network models used were effective in detecting the quality of surgery performed in patients with unilateral cleft lips, with the MobileNet and EfficientNet models having the best results with values above 80% in all metrics. Conclusions. The artificial intelligence models used have demonstrated their ability to measure postoperative lip symmetry in patients operated on for unilateral cleft lip. The best models were MobileNet and EfficientNet with metrics above 80%. This efficiency can be improved by increasing the number of data.

Instituciones: Universidad Peruana de Ciencias Aplicadas; Universidad Peruana Unión