Non-Neural 3D Nasal Reconstruction: A Sparse Landmark Algorithmic Approach for Medical Applications

This paper presents a novel method for reconstructing a highly accurate 3D nose model of the human from 2D images and pre-marked landmarks based on algorithmic methods. The study focuses on the reconstruction of a 3D nose model tailored for applications in healthcare and cosmetic surgery. The approach leverages advanced image processing techniques, 3D Morphable Models (3DMM), and deformation techniques to overcome the limitations of deep learning models, particularly addressing the interpretability issues commonly encountered in medical applications. The proposed method estimates the 3D coordinates of landmark points using a 3D structure estimation algorithm. Sub-landmarks are extracted through image processing techniques and interpolation. The initial surface is generated using a 3DMM, though its accuracy remains limited. To enhance precision, deformation techniques are applied, utilizing the coordinates of 76 identified landmarks and sub-landmarks. The resulting 3D nose model is constructed based on algorithmic methods and pre-marked landmarks. Evaluation of the 3D model is conducted by comparing landmark distances and shape similarity with expert-determined ground truth on 30 Vietnamese volunteers aged 18 to 47, all of whom were either preparing for or required nasal surgery. Experimental results demonstrate a strong agreement between the reconstructed 3D model and the ground truth. The method achieved a mean landmark distance error of 0.631 mm and a shape error of 1.738 mm, demonstrating its potential for medical applications.