Page 129 - IJB-7-4

Li, et al
agencies, local communities, and individual makers have To the best of authors’ knowledge, there has not been a
used AM as a supply chain response to de-centralize study demonstrating a completely automated process for
production and combat supply chain disruption for RPE customizing respirator masks and validated against a 3D
and other protective equipment [48-57] . face dataset.
Despite the promises of AM in producing custom Recently, the authors proposed a new automated
[65]
fit RPE, major cost barriers exist in adopting AM for respirator mask customization process which reduced
large-scale production. One key cost contributor stems design time from hours to minutes. While it is a promising
from a highly manual and time-consuming customization pipeline, the pipeline has yet to be seamlessly integrated,
process, which can add up to 20 – 30% of the overall neither has it been validated against a reasonable dataset.
[65]
AM production cost [25,34,37,58] . Sporadic efforts have been This study builds upon this previous work and
made over the past few decades to simplify and shorten integrates a CAD Application Programming Interface
the AM design process. The most utilized method has (API) with the rest of the pipeline to present a seamless
been parametric geometric modeling, where a generic and automated design pipeline for generating custom-
parametric computer-aided design (CAD) model is fitted respirator masks ready to be manufactured using
created with control points that can be automatically AM techniques.
updated based on the shape of an input scan. Such To investigate whether this pipeline can be
approach has been applied to customizable medical universally applied to people from different demographic
devices, such as protective face masks, wrist splits, backgrounds, an online portal was created to recruit
ankle-foot orthotics [59-61] , and recently to BIPAP/CPAP participants during the COVID-19 lockdown periods
masks [38,40] . However, these are semi-automated modeling in the UK, where their facial scans and demographic
processes where a technician with CAD modeling information were collected. Success rate, computational
knowledge is needed to perform manual operations such run time, and fit (how well a mask fits to a face) were
as aligning the generic model to a raw scan. Studies evaluated. Furthermore, fit results were compared across
demonstrating a fully automated customization process subcategories of demographics to investigate whether
for traditionally mass-produced body-fitted products the pipeline can produce respirator mask models that fit
are rare. Ellena et al. proposed a design process for equally well to people across different age (young, middle
[62]
customizing bicycle helmets, where a statistical shape aged, senior), gender (male or female), ethnic (Asian,
model was utilized to classify a head scan into one out of White, and Others), or BMI groups (healthy, overweight,
four helmet sizes before cropping away the inner lining of and obese).
the helmet with a generic B-Spline head surface adapted 2. Methods
to the shape of the head scan using an iterative genetic
algorithm . While it is a fully automated process,
[63]
it essentially uses a sizing-based approach where the 2.1. RPE design pipeline
statistical shape model was built by analyzing 222 head An automated MC design pipeline, shown in Figure 1,
scans that may be representative of the Australian cyclist was employed for this study. This pipeline is based on our
population, but not other demographic groups. For RPE previous study which employs a series of alignment and
designs, the sample size required could be much larger template fitting processes to represent user-submitted
[65]
(e.g. the US NIOSH panel used almost 4000 subjects) 3D facial scans using a universal 3D face template
for a particular demographic group. The amount of mesh. This removes heterogeneity across different raw
resources and time needed to collect a large-scale 3D facial meshes in terms of orientation, location, and
anthropometric database and to build a statistical shape mesh structure (vertex indexing and triangulation),
model for each applicable demographic group defeats the thereby enabling the subsequent automatic extraction of
purpose of lowering design cost for tailor-fit products. topographical data from a large facial dataset. Using a set
Sela et al. proposed a fully automated pipeline to of predefined vertices on the template mesh as landmarks,
[64]
generate customized CPAP masks based on a 3D scan or 200 points on two egg-shaped loops based on those
depth image of a person’s face. The shape of the mask landmark locations were projected onto the fitted template
model was customized by updating 256 control points on a
generic mask model made up with Non-uniform Rational
Basis Spline (NURBS) surfaces. However, automatically
deforming a generic NURBS-based geometric model
with organic shape can be problematic, particularly given
the complexity and high variability of facial shapes. The Figure 1. Pipeline overview, the computer-aided design model
pipeline was only validated against one subject; therefore, generation step (highlighted) is now integrated into the design
its robustness against a larger dataset remains unknown. pipeline.

International Journal of Bioprinting (2021)–Volume 7, Issue 4 125