Novel Baseline Facial Muscle Database Using Statistical Shape Modeling and In Silico Trials toward Decision Support for Facial Rehabilitation

Backgrounds and Objective: Facial palsy is a complex pathophysiological condition affecting the personal and professional lives of the involved patients. Sudden muscle weakness or paralysis needs to be rehabilitated to recover a symmetric and expressive face. Computer-aided decision support systems for facial rehabilitation have been developed. However, there is a lack of facial muscle baseline data to evaluate the patient states and guide as well as optimize the rehabilitation strategy. In this present study, we aimed to develop a novel baseline facial muscle database (static and dynamic behaviors) using the coupling between statistical shape modeling and in-silico trial approaches. Methods: 10,000 virtual subjects (5000 males and 5000 females) were generated from a statistical shape modeling (SSM) head model. Skull and muscle networks were defined so that they statistically fit with the head shapes. Two standard mimics: smiling and kissing were generated. The muscle strains of the lengths in neutral and mimic positions were computed and recorded thanks to the muscle insertion and attachment points on the animated head and skull meshes. For validation, five head and skull meshes were reconstructed from the five computed tomography (CT) image sets. Skull and muscle networks were then predicted from the reconstructed head meshes. The predicted skull meshes were compared with the reconstructed skull meshes based on the mesh-to-mesh distance metrics. The predicted muscle lengths were also compared with those manually defined on the reconstructed head and skull meshes. Moreover, the computed muscle lengths and strains were compared with those in our previous studies and the literature. Results: The skull prediction's median deviations from the CT-based models were 2.2236 mm, 2.1371 mm, and 2.1277 mm for the skull shape, skull mesh, and muscle attachment point regions, respectively. The median deviation of the muscle lengths was 4.8940 mm. The computed muscle strains were compatible with the reported values in our previous Kinect-based method and the literature. Conclusions: The development of our novel facial muscle database opens new avenues to accurately evaluate the facial muscle states of facial palsy patients. Based on the evaluated results, specific types of facial mimic rehabilitation exercises can also be selected optimally to train the target muscles. In perspective, the database of the computed muscle lengths and strains will be integrated into our available clinical decision support system for automatically detecting malfunctioning muscles and proposing patient-specific rehabilitation serious games.

Reinforcement learning coupled with finite element modeling for facial motion learning

BACKGROUND AND OBJECTIVE

Facial palsy patients or patients with facial transplantation have abnormal facial motion due to altered facial muscle functions and nerve damage. Computer-aided system and physics-based models have been developed to provide objective and quantitative information. However, the predictive capacity of these solutions is still limited to explore the facial motion patterns with emerging properties. The present study aims to couple the reinforcement learning and the finite element modeling for facial motion learning and prediction.

METHODS

A novel modeling workflow for learning facial motion was developed. A physically-based model of the face within the Artisynth modeling platform was used. Information exchange protocol was proposed to link reinforcement learning and rigid multi-bodies dynamics outcomes. Two reinforcement learning algorithms (deep deterministic policy gradient (DDPG) and Twin-delayed DDPG (TD3)) were used and implemented to drive the simulations of symmetry-oriented and smile movements. Numerical outcomes were compared to experimental observations (Bosphorus database) for evaluation and validation purposes.

RESULTS

As result, after more than 100 episodes of exploring the environment, the agent starts to learn from previous trials and can find the optimal policy after more than 300 episodes of training. Regarding the symmetry-oriented motion, the muscle excitations predicted by the trained agent help to increase the value of reward from R = -2.06 to R = -0.23, which counts for ∼89% improvement of the symmetry value of the face. For smile-oriented motion, two points at the edge of the mouth move up 0.35 cm, which is within the range of movements estimated from the Bosphorus database (0.4 ± 0.32 cm).

CONCLUSIONS

The present study explored the muscle excitation patterns by coupling reinforcement learning with a detailed finite element model of the face. We developed, for the first time, a novel coupling scheme to integrate the finite element simulation into the reinforcement learning process for facial motion learning. As perspectives, this present workflow will be applied for facial palsy and facial transplantation patients to guide and optimize the functional rehabilitation program.

Enhanced head-skull shape learning using statistical modeling and topological features

Skull prediction from the head is a challenging issue toward a cost-effective therapeutic solution for facial disorders. This issue was initially studied in our previous work using full head-to-skull relationship learning. However, the head-skull thickness topology is locally shaped, especially in the face region. Thus, the objective of the present study was to enhance our head-to-skull prediction problem by using local topological features for training and predicting. Head and skull feature points were sampled on 329 head and skull models from computed tomography (CT) images. These feature points were classified into the back and facial topologies. Head-to-skull relations were trained using the partial least square regression (PLSR) models separately in the two topologies. A hyperparameter tuning process was also conducted for selecting optimal parameters for each training model. Thus, a new skull could be generated so that its shape was statistically fitted with the target head. Mean errors of the predicted skulls using the topology-based learning method were better than those using the non-topology-based learning method. After tenfold cross-validation, the mean error was enhanced 36.96% for the skull shapes and 14.17% for the skull models. Mean error in the facial skull region was especially improved with 4.98%. The mean errors were also improved 11.71% and 25.74% in the muscle attachment regions and the back skull regions respectively. Moreover, using the enhanced learning strategy, the errors (mean ± SD) for the best and worst prediction cases are from 1.1994 ± 1.1225 mm (median: 0.9036, coefficient of multiple determination (R²): 0.997274) to 3.6972 ± 2.4118 mm (median: 3.9089, R²: 0.999614) and from 2.0172 ± 2.0454 mm (median: 1.2999, R²: 0.995959) to 4.0227 ± 2.6098 mm (median: 3.9998, R²: 0.998577) for the predicted skull shapes and the predicted skull models respectively. This present study showed that more detailed information on the head-skull shape leads to a better accuracy level for the skull prediction from the head. In particular, local topological features on the back and face regions of interest should be considered toward a better learning strategy for the head-to-skull prediction problem. In perspective, this enhanced learning strategy was used to update our developed clinical decision support system for facial disorders. Furthermore, a new class of learning methods, called geometric deep learning will be studied.

Kinect-driven Patient-specific Head, Skull, and Muscle Network Modelling for Facial Palsy Patients

BACKGROUND AND OBJECTIVE

Facial palsy negatively affects both professional and personal life qualities of involved patients. Classical facial rehabilitation strategies can recover facial mimics into their normal and symmetrical movements and appearances. However, there is a lack of objective, quantitative, and in-vivo facial texture and muscle activation bio-feedbacks for personalizing rehabilitation programs and diagnosing recovering progresses. Consequently, this study proposed a novel patient-specific modelling method for generating a full patient specific head model from a visual sensor and then computing the facial texture and muscle activation in real-time for further clinical decision making.

METHODS

The modeling workflow includes (1) Kinect-to-head, (2) head-to-skull, and (3) muscle network definition & generation processes. In the Kinect-to-head process, subject-specific data acquired from a new user in neutral mimic were used for generating his/her geometrical head model with facial texture. In particular, a template head model was deformed to optimally fit with high-definition facial points acquired by the Kinect sensor. Moreover, the facial texture was also merged from his/her facial images in left, right, and center points of view. In the head-to-skull process, a generic skull model was deformed so that its shape was statistically fitted with his/her geometrical head model. In the muscle network definition & generation process, a muscle network was defined from the head and skull models for computing muscle strains during facial movements. Muscle insertion points and muscle attachment points were defined as vertex positions on the head model and the skull model respectively based on the standard facial anatomy. Three healthy subjects and two facial palsy patients were selected for validating the proposed method. In neutral positions, magnetic resonance imaging (MRI)-based head and skull models were compared with Kinect-based head and skull models. In mimic positions, infrared depth-based head models in smiling and [u]-pronouncing mimics were compared with appropriate animated Kinect-driven head models. The Hausdorff distance metric was used for these comparisons. Moreover, computed muscle lengths and strains in the tested facial mimics were validated with reported values in literature.

RESULTS

With the current hardware configuration, the patient-specific head model with skull and muscle network could be fast generated within 17.16±0.37s and animated in real-time with the framerate of 40 fps. In neutral positions, the best mean error was 1.91 mm for the head models and 3.21 mm for the skull models. On facial regions, the best mean errors were 1.53 mm and 2.82 mm for head and skull models respectively. On muscle insertion/attachment point regions, the best mean errors were 1.09 mm and 2.16 mm for head and skull models respectively. In mimic positions, these errors were 2.02 mm in smiling mimics and 2.00 mm in [u]-pronouncing mimics for the head models on facial regions. All above error values were computed on a one-time validation procedure. Facial muscles exhibited muscle shortening and muscle elongating for smiling and pronunciation of sound [u] respectively. Extracted muscle features (i.e. muscle length and strain) are in agreement with experimental and literature data.

CONCLUSIONS

This study proposed a novel modeling method for fast generating and animating patient-specific biomechanical head model with facial texture and muscle activation bio-feedbacks. The Kinect-driven muscle strains could be applied for further real-time muscle-oriented facial paralysis grading and other facial analysis applications.

A Systematic Review of Real-Time Medical Simulations with Soft-Tissue Deformation: Computational Approaches, Interaction Devices, System Architectures, and Clinical Validations

Simulating deformations of soft tissues is a complex engineering task, and it is even more difficult when facing the constraint between computation speed and system accuracy. However, literature lacks of a holistic review of all necessary aspects (computational approaches, interaction devices, system architectures, and clinical validations) for developing an effective system of soft-tissue simulations. This paper summarizes and analyses recent achievements of resolving these issues to estimate general trends and weakness for future developments. A systematic review process was conducted using the PRISMA protocol with three reliable scientific search engines (ScienceDirect, PubMed, and IEEE). Fifty-five relevant papers were finally selected and included into the review process, and a quality assessment procedure was also performed on them. The computational approaches were categorized into mesh, meshfree, and hybrid approaches. The interaction devices concerned about combination between virtual surgical instruments and force-feedback devices, 3D scanners, biomechanical sensors, human interface devices, 3D viewers, and 2D/3D optical cameras. System architectures were analysed based on the concepts of system execution schemes and system frameworks. In particular, system execution schemes included distribution-based, multithread-based, and multimodel-based executions. System frameworks are grouped into the input and output interaction frameworks, the graphic interaction frameworks, the modelling frameworks, and the hybrid frameworks. Clinical validation procedures are ordered as three levels: geometrical validation, model behavior validation, and user acceptability/safety validation. The present review paper provides useful information to characterize how real-time medical simulation systems with soft-tissue deformations have been developed. By clearly analysing advantages and drawbacks in each system development aspect, this review can be used as a reference guideline for developing systems of soft-tissue simulations.

Real-time Subject-specific Head and Facial Mimic Animation System using a Contactless Kinect Sensor and System of Systems Approach

Facial palsies due to stroke, accidental and sportive injuries or sometimes without etiology, affect the professional and personal lives of involved patients. These disorders are not only a functional handicap but also a social integration impairment. The recovery of facial mimics with a normal and symmetrical facial expression allows involved patients to improve their living conditions and social identity. Current approaches lack of visual feedbacks. To monitor facial mimics and head movements in a quantitative and objective manners, a computer-aided animation system needs to be developed. Numerous systems have been proposed using single camera, stereo camera, 3-D scanner, and Kinect approaches. In particular, Kinect contactless sensor has been proven to be very suitable for 3-D facial simulation applications. However, little studies have employed the Kinect sensor for real-time head animation applications. Consequently, this study developed a real-time head and facial mimic animation system using the contactless Kinect sensor and the system of systems approach. To evaluate the accuracy of the subject-specific Kinect-based geometrical models, magnetic resonance imaging (MRI) data were used. As results, the mean distance deviation between generated Kinect-based and reconstructed MRI-based geometrical head models are approximately 1 mm for two tested subjects. The generation times are 9.7 s ± 0.3 and 0.046 s ± 0.005 by using the full facial landmarks and MPEG-4 facial landmarks respectively. Real-time head and facial mimic animations were illustrated. Particularly, the system could be executed at a very high framerate (60 fps). Further developments relate to the integration of texture information and internal structures such as a skull and muscle network to develop a full subject specific head and facial mimic animation system for facial mimic rehabilitation.