Convolutional Neural Network Models for Automatic Preoperative Severity Assessment in Unilateral Cleft Lip.

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Craniofacial anomaly assessment

Background:

• Lack of standardized scales for preoperative cleft lip severity.
• Machine learning (ML) has not been applied to cleft lip disease classification.
• Need for objective and automated methods for assessing cleft lip morphology.

Purpose of the Study:

• To develop and evaluate ML models for automated detection and measurement of facial landmarks in cleft lip patients.
• To assess the feasibility of using ML for assigning preoperative severity grades for cleft lip.
• To explore the potential of ML in standardizing cleft lip morphology classification.

Main Methods:

• Trained five convolutional neural network (CNN) models on 800 preoperative unilateral cleft lip images.
• Manually annotated images for cleft-specific landmarks and used expert ratings for severity.
• Calculated mean squared error and Pearson correlation for cleft width, nostril width, and severity grade assignment.

Main Results:

• All five CNN models demonstrated good performance in landmark detection and severity grading.
• The Residual Network model achieved the highest accuracy (severity correlation: 0.892).
• MobileNet showed high accuracy (severity correlation: 0.860) and is compatible with mobile devices.

Conclusions:

• ML models can accurately measure facial features and assign cleft lip severity grades.
• Automated ML approaches offer a promising solution for classifying cleft lip morphology.
• Potential for mobile applications to provide real-time clinical decision support and patient counseling.