It’s Time to Revamp Mask Fitting
Part 2: Will AI Save the Game?
In my previous post, I introduced the concepts of fit test for masks — specifically N95 respirators, including Qualitative Fit Test (QLFT) and Quantitative Fit Test (QNFT). I then discussed the current nuances and limitations of each procedure including time, cost of test kits, requirement of trained crew, false positives (QLFT), and damage to the masks (QNFT). Finally, I briefly mentioned AI as a potential game changer in improving current practices. This point shall be the focus of this post.
Status of AI in mask fitting
To date, little research has been carried out to investigate how AI could be utilized for mask fitting. The most common studies combining AI and masks are the ones that use neural networks to tell if a person is wearing a mask at all. These studies typically employ similar technology as used in facial recognition to detect the existence of masks on a photo or video. While sharing the same keywords, they are quite distant from the interest of this article — using AI to improve the fitting processing for N95 respirators.
The earliest work implementing AI, in particular Computer Vision on 3D scanning for mask fitting, was conducted by Lui et al from City University of Hong Kong in 2011. The study consists of two parts: it investigates the use of machine learning algorithms for developing classification predictors of how well a respirator fits a wearer, and how low-cost 3D scanning technology based on structured light could obtain the facial anthropometric data used for the prediction. The authors conclude that a predictor system to augment the fit testing procedure can be both feasible and useful. Strangely, this rather foresightful work has not been followed up by researchers in related fields, despite the rapid development of neural networks between 2012 and 2021.
More recent research of considerable interest was conducted by Biagiotti et al in 2019, where the authors devised a mask analysis and size quantification framework, utilizing 2D images of the test subject to predict respirator size and fit. This work provides the unique insight that 2D images can be reconstructed into 3D for our mask fitting problem.
Apart from these scattered works, it would be safe for us to assume that there is no large scale or coherent effort in academia or industry to interrogate the potential for AI-assisted mask fitting. Before stepping into the technicalities, it is helpful to explore the unique challenges faced by mask fitting that may impede successful development of new methods.
Challenges of using AI
First, mask fitting is prone to a “butterfly effect” of the user’s facial features. A Cambridge study in 2020 showed us that a minor difference in facial dimensions or subcutaneous fat — a matter of 3–4 mm difference — could lead to significant quantitative fit differences. As an example, two of the tested participants were mother and daughter with highly similar facial features, yet the fit factor scores achieved by the two individuals bore no correlation. This indicates that even if the same person presents with minor changes in weight, the fit of the mask may need to be reassessed. This “butterfly effect” certainly places importance on the accuracy of both 3D scanning techniques, and the deep learning models applied to the 3D image.
Second, “minor” facial features such as facial hair may significantly disrupt the fitting assessment. Conventional wisdom tells us that facial hair may help trap the particles in the air — but this is not true. Research shows that the presence of facial hair under the sealing surface causes 20 to 1000 times more leakage compared to clean-shaven individuals. In fact, some studies have shown that even a day or two of stubble can begin to reduce protection. While the CDC requires medical professionals to trim facial hairs to styles that offer the best seal, if not clean shaven, there is no guarantee that facial hairs will not be present in the game of mask fitting, presenting a non-negligible challenge to our initiative.
Are we ready?
So, are we ready for to welcome AI into the mask-fitting game? There are a few components of the technological requirements we can analyze. The first is 3D scanning. For it to make operational sense, the 3D scanning of a participant’s face needs to be finished within (or better, far ahead of) the typical timeframe of current mask fitting procedures (15–20 mins). A few commercially available technologies we can find online gives us an estimate of aproximately 1 minute of scanning time and 3–4 minutes of processing time to generate a high-quality 3D head scan. The high efficiency of existing technologies means that this step is unlikely to be the bottleneck of the whole AI-guided mask fitting process.
Next is the centerpiece — machine learning on 3D images(to avoid confusion, we denote the scans, no matter 2D or 3D, as “images” instead of “models”, to differentiate from machine Llearning models). This is a heavily researched topic due to its tremendous potential for medical imaging such as X-ray, CT, and MRI. In recent years, there has been an exponential increase in the application of 3D deep learning on medical image segmentation, classification, detection, and localization. A review paper by researchers from Nanyang Technological University, Singapore and Karolinska Institute, Sweden collected 132 papers on these topics.
Thanks to breakthroughs in deep learning models such as VGGNet, GoogleLeNet, and ResNet, and their 3D counterparts, training deep learning models on 3D medical images is now impossible. Several works on (detecting?) brain tumor/lesion segmentation in CT scans give us ~80% precision levels. Greater challenges, however, lie in the stages datasets suitable for training can be collected.
Pieces of the Puzzle
First is the well-known need for massive datasets (for familiar applications such as facial recognition, training setscan reach up to a million data points). For our mask fitting project, it is unlikely that researchers will be able to acquire millions of 3D scans of participants, for both technical and regulatory reasons. This means that techniques to reduce the input dataset size requirement, such as transfer learning, will likely be needed.
Second, certain pre-processing steps on the 3D images may be needed before they are fed into deep learning models. Facial scans of participants taken by different devices and circumstances are likely to be of different qualities and may need to be normalized before training. Facial hairs, as previously mentioned, may prove to be a nuisance for both 3D image generation and model training. All these steps will result in an extra workload for researchers.
Third, for AI-assisted mask fitting to make practical sense the accuracy of the model predictions must be on par with, or better than human level results. Researchers will certainly need to find a way to fill the considerable gap between model and human precision that still exists in many areas of 3D medical imaging.
And last, care needs to be taken for the training data to encompass the diversity and multitude of facial features. For an AI-assisted mask fitting procedure to be successful, it must work well on medical practitioners of any demographic. Regarding facial recognition, for instance, a NIST study in 2019 found “empirical evidence for the existence of demographic differentials in the majority of the face recognition algorithms we studied.”
To correct the bias, we ought not go down the slippery path of direct human intervention on model parameters. As the study also points out, “not all algorithms give this high rate of false positives across demographics in one-to-many matching, and those that are the most equitable also rank among the most accurate.” We believe this to be a technical challenge that can be solved by methods such as more balanced sample selection.
To conclude, here are some thoughts about AI-assisted N95 respirator fitting, and potential ways to improve it: the current procedures have limitations such as time, expertise requirements, price of test kits, and damage of masks. AI could potentially be the game changer of this landscape. The panics of the pandemic will hopefully one day fade from our memory, but we should not let go of the opportunities it brought to revolutionize our healthcare practices. The quest is present. Are we ready to undertake it?
