A Deep Learning Approach for Predicting Subject-Specific Human Skull Shape from Head Toward a Decision Support System for Home-Based Facial Rehabilitation

Objective

Prediction of human skull shape from head is a complex and challenging engineering task for the development of a computer-aided vision system. Skull-to-face generation has been commonly performed in forensic facial reconstruction. Classical statistical approaches were usually used. However, the head-to-skull relationship is still misunderstood. Recently, novel deep learning (DL) models have showed their efficiency and robustness for a large range of applications. The present study aimed to develop a novel approach based on deep learning models to reconstruct the human skull shape from head.

Material and methods

A head-to-skull generation workflow was developed and evaluated. A database of computed tomography (CT) images of 209 subjects was established for training and testing purposes. Three-dimension (3-D) head and skull geometries were reconstructed and then their respective descriptors (head/skull volumes, sampling feature points and point-to-center distances, head-skull thickness, Gaussian curvatures) were extracted. Two deep learning models (regression neural network and long-short term memory (LSTM)) were implemented and evaluated with different learning configurations. A 10-fold cross-validation was performed. Finally, the best and worst predicted cases were analyzed and discussed.

Results

The mean errors from 10-fold cross-validation showed a better accuracy level for the regression neural network model according to the long short-term memory model. The mean error between the DL-predicted skull shapes and CT-based skull shapes ranges from 1.67 mm to 3.99 mm by using the regression deep learning model and the best learning configuration. The volume deviation between predicted skull shapes and CT-based skull shapes is smaller than 5%.

Conclusions

The present study suggested that regression deep learning model allows human skull to be predicted from a given head with a good level of accuracy. This opens new avenues for the rapid generation of human skull shape from visual sensors (e.g. Microsoft Kinect) toward a computer-aided vision system for facial mimic rehabilitation. As perspectives, muscle network will be incorporated into the present workflow. Then, facial mimic movements will be tracked and animated to evaluate and optimize the rehabilitation movements and exercises.