Evaluation of reconstructed auricles by convolutional neural networks

Area of Science:

• Plastic Surgery
• Artificial Intelligence
• Medical Imaging

Background:

• Determining crucial structures for natural-looking ear reconstruction remains a challenge for surgeons.
• Objective evaluation of auricle aesthetics is essential for successful reconstructive surgery.

Purpose of the Study:

• To explore the feasibility of using convolutional neural network (CNN) models to evaluate reconstructed auricles.
• To establish criteria for natural-looking auricle design by visualizing CNN model attention.

Main Methods:

• Trained ResNet CNN models on 20 labeled datasets (400 images) pre-trained on ImageNet.
• Calculated saliency maps and occlusion maps to analyze model attention.
• Correlated CNN model evaluations with those of 20 volunteers and medical students.

Main Results:

• CNN models achieved an average accuracy of 0.8245 ± 0.0356, exceeding the 0.80 threshold.
• CNN model evaluations showed a significant correlation with medical student assessments (P < 0.05).
• Attention visualization highlighted key aesthetic features: helix contour, antihelix inferior crura, and concha contour.

Conclusions:

• CNNs can effectively evaluate reconstructed auricles, mimicking human expert judgment.
• Attention maps from CNNs provide insights into subjective aesthetic criteria for auricle reconstruction.
• This approach aids in defining objective standards for natural-looking ear aesthetics.