Deep Facial Landmark Factorization on 3d Craniofacial Asymmetry and Development of a Scalable Platform

Scope and Significance. Facial symmetry is an important clinical factor related to appearance attractiveness. Perfect facial symmetry is considered merely a theoretical existence; nonetheless, significant fluctuate asymmetry can cause aesthetic or even functional problems. Qualitative elaboration of facial asymmetry can be ambiguous if without quantitative support. Further, perception and assessments of asymmetry can be obscured by shades, occlusion, skin textures, etc. Unbalanced and opaque information with which physicians and patients communicate and anticipation gaps between the two parties usually lead to medical disputes. Whether facial asymmetry is congenital, developmental, or acquired, effective assessment and treatment planning and management are crucial topics between the stakeholders and need to be well addressed. Objectives. Following the pilot study of our on-going MOST project, the aims of the proposed project are threefold: (1) to quantify facial asymmetry in terms of the proposed overall asymmetry index (oAI); (2) to construct a facial asymmetry classifier that can morphometric ally categorizes craniofacial images as either one of the three classes: perceived normal (PN), perceived asymmetrically normal (PAN), and perceived abnormal (PA), that can adapt itself with ever-fed subjective vs. quantitative classifications across different age groups, and that is scalable and deep enough to factorize and learn facial landmarks; (3) to elaborate the confidence index (Ci) of the classification, interpret the phenomenon of “ambiguous asymmetry perception,” and present the physician-patient mutually transparent trio-information–{oAI,PN|PAN|PA,Ci}. The consistency between the trio-information {oAI,PN|PAN|PA,Ci} and clinical readings can be considered as a key performance factor (KPI) of the proposed project. Study Design and Schematic Procedures. (1) Acquire a set of cone-beam computed tomography (CBCT) and 3dMD images from a normal face; (2) Extract facial surface which will be then mono-chromed and the skin texture removed to alleviate asymmetry classification from being interfered by appearance and skin quality; (3) Identify facial medial and bilateral 3D facial landmarks to form facial features; (4) Morph the non-textured face into sixty-four or more “stimulus faces” with varied degree of asymmetry; chin and nose rotations/displacements will be employed to achieve transformation; (5) Conducted visual questionnaire surveys and categorize these sixty-four or more faces into either PN, PAN, or PA, along with corresponding Ci, respectively; (6) Construct the asymmetry classifier that can be trained to predict and that can be adaptive to scalable to implement clinical validation and factorize the facial landmarks; (7) Compute the adaptive trio-information as a tool conveyed in clinical practice. The steps (5), (6), and (7) are expected to be inter-correlated. Academic and Clinical Contributions. This study will be a pioneering quantitative, theoretical and thorough study of facial asymmetry. We shall build and then keep revising a software platform that accommodates the algorithms and interactive functionalities addressed during the span of the proposed project. The physician and the patient can communicate over the presented trio-information of the patient’s face and determine whether go or no-go with respect to the possible treatments that follow in a candid and transparent fashion. Further, the proposed asymmetry classifier shall evolve from a primitive back-propagation neural work classifier to a model that possesses deep learning capability.

Project ID:PB10607-1402
External Project ID:MOST106-2221-E182-026