Open Access

ARTICLE

A Novel Face-to-Skull Prediction Based on Face-to-Back Head Relation

Tien-Tuan Dao¹, Lan-Nhi Tran-Ngoc^2,3, Trong-Pham Nguyen-Huu^2,3, Khanh-Linh Dinh-Bui^2,3, Nhat-Minh Nguyen^2,3, Ngoc-Bich Le^2,3, Tan-Nhu Nguyen^2,3,*

1 Univ. Lille, CNRS, Centrale Lille, UMR 9013-LaMcube-Laboratoire de Mécanique, Multiphysique, Multiéchelle, F-59000, Lille, France
2 School of Biomedical Engineering, International University, Ho Chi Minh City, 700000, Vietnam
3 Vietnam National University Ho Chi Minh City, Vietnam

* Corresponding Author: Tan-Nhu Nguyen. Email:

(This article belongs to the Special Issue: Beyond the Surface: Exploring the Depths of Deep Learning in Face Recognition)

Computers, Materials & Continua 2025, 84(2), 3345-3369. https://doi.org/10.32604/cmc.2025.065279

Received 08 March 2025; Accepted 20 May 2025; Issue published 03 July 2025

Abstract

Skull structures are important for biomechanical head simulations, but they are mostly reconstructed from medical images. These reconstruction methods harm the human body and have a long processing time. Currently, skull structures can be straightforwardly predicted from the head, but a full head shape must be available. Most scanning devices can only capture the face shape. Consequently, a method that can quickly predict the full skull structures from the face is necessary. In this study, a novel face-to-skull prediction procedure is introduced. Given a three-dimensional (3-D) face shape, a skull mesh could be predicted so that its shape would statistically fit the face shape. Several prediction strategies were conducted. The optimal prediction strategy with its optimal hyperparameters was experimentally selected through a ten-fold cross-validation with 329 subjects. As a result, the face-to-skull prediction strategy based on the relations between face head shape and back head shape, between face head shape and face skull shape, and between back head shape and back skull shape was optimal. The optimal mean mesh-to-mesh distance (mean ± SD) between the predicted skull shapes and the ground truth skull shapes was 1.93 ± 0.36 mm, and those between the predicted skull meshes and the ground truth skull meshes were 2.65 ± 0.36 mm. Moreover, the prediction errors in back-skull and muscle attachment regions were 1.7432 ± 0.5217 mm and 1.7671 ± 0.3829 mm, respectively. These errors are within the acceptable range of facial muscle simulation. In perspective, this method will be employed in our clinical decision support system to enhance the accuracy of biomechanical head simulation based on a stereo fusion camera system. Moreover, we will also enhance the accuracy of the face-to-skull prediction by diversifying the dataset into more varied geographical regions and genders. More types of parameters, such as Body Mass Index (BMI), coupled with head-to-skull thicknesses, will be fused with the proposed face-to-skull procedure.

---

Keywords

---