Application of convolutional neural networks for nasofacial reconstruction

The nasal prosthesis is an artificial solution to the restoration of tissue in the nasal region, traditionally requiring time-intensive manual sculpting directly on the patient's face. Whether a novel, automated approach using Convolutional Neural Networks (CNNs) can reconstruct the nasofacial region from 2-dimensional (2D) photography is unclear.

The purpose of this study was to evaluate whether the automated approach could generate 3-dimensional (3D) models suitable for direct 3D printing.

The Residual U-Net architecture was trained on a dataset derived from CelebA-HQ comprising 12 000 image pairs (input images with trapezoidal masks and corresponding mask-free output images). The dataset's normalized and square-aligned images facilitated efficient training, with the masked area prioritized during reconstruction through weighted loss functions. The model's performance was evaluated using Peak Signal-to-Noise Ratio (PSNR).

Accurate reconstruction of nasal geometries that harmonized well with the overall facial structure were demonstrated. Visual assessments confirmed the approach's effectiveness, although minor limitations such as inconsistent texture modeling were observed under nonideal conditions. Beyond technical validation, this automated solution should reduce laboratory and clinical time requirements, minimizing patient travel and emotional stress and enhance accessibility to personalized prosthetic care. By leveraging advancements in CNNs and image-to-3D model conversion and manufacturing techniques, this approach offered a suitable path toward automating the creation of nasal prostheses.

This work represents a significant step toward automating the creation of nasal prostheses, enabling faster, more cost-effective, and patient-friendly solutions for patients with midfacial deficiency.