"PROJECT DESCRIPTIONAims and ObjectivesThe diagnosis of Obstructive Sleep Apnoea (OSA), the most prevalent sleeping disorder,presents a significant problem to health care systems worldwide.Historically OSA diagnosishas been made with laboratory-based polysomnography (PSG) studies which are very time- consuming and expensive.Difficulties related to accessing PSG studies and the large numberof individuals in the community who are thought to be undiagnosed has led to a substantialinternational effort to find alternative, simpler methods to detect OSA. The overriding aim ofthis project is to develop algorithms for a quick, safe and cost-effective way of predictingOSA for the early prevention of the syndrome. Computer Vision-based algorithms will bedeveloped to automatically extract OSA phenotypic features (already known in the literature)from three dimensional (3D) photographs of the face and neck areas of a person. Thesefeatures along with other 2D and 3D image features recently demonstrated to be effective inthe area of object detection (such as local descriptors and histogram of gradients) will then bepassed to different a state of the art machine learning algorithms (including deep learning) fortraining. Exhaustive training on a large number of normal and sleep apnoeic samples wouldenable us to select the most distinctive features to be used for the prediction of moderate tosevere levels of OSA.Specific aims of the project are to use surface and volumetric images of the head andneck to:Aim-1 Develop algorithms to extract automatically 2D and 3D craniofacialfeatures that potentially phenotype OSA.Aim-2Develop feature selection and classification algorithms to predict OSA.BackgroundMany attempts have been made in the past to predict OSA based on questionnaires. Forexample, the Berlin questionnaire predicts the level of risk based on snoring, tiredness, bloodpressure and body mass index information while the Epworth Sleepiness questionnaireassesses the sleepiness in various situations during theday. Although they are self-administered and low-cost,they have shortcomings in accurately identifying9affected individuals. A systematic review of eightdifferent questionnaire models shows substantialvariation in the diagnostic performance among them,11 none presenting reasonable sensitivity and specificity.Imaging techniques have been considered asFigure 1—ROC curves of Logistic Regressionuseful adjunctive tools to diagnose OSA, with theModels for predicting OSA with 2Dphotographic measurements.radiographic head film (cephalometric) analysis being12 the most convenient and widely used. This technique has been used determining significant13 13correlation between OSA severity and neck circumference in men, retropalatal airway, 14,15 15,16 17Upper Airway Length (UAL), inferiorly placed hyoid bone, facial depth, mandibular17 16 plane angle and reduced mid-face length. Cone Beam Computed Tomography (CBCT),Medical CT, Magnetic Resonance Imaging are three dimensional counterpart ofcephalometric images and hence provide more appealing results but they are moreinvasive19and expensive.Recently, Lee et al. have explored that digital photographs of craniofacialsurface structures may be used to predict OSA. They analysed frontal and profile photographsof 114 subjects. Using only four photographic measurements (face width, eye width,cervicomental angle and mandibular length) their logistic regression based model obtained anaccuracy of 76.1% with an area under Receiver Operating Characteristic (ROC) curve of0.82. Their results also show these features capturing the composite elements of craniofacial structures and regional adiposity can predict OSA better than demographic data (e.g. BMI orneck circumference) collected by clinical observations. However, like cephalometry, digitalphotographs are two dimensional in nature and hence neither non-linear measurements normeasurements of the shape of craniofacial anatomy can be obtained.Three dimensional surface imaging technologies have recently been developed thatare well suited for imaging the human head, face and neck.Images can be accuratelyobtained extremely quickly (less than one second) and the technique is non-invasive in natureas it does not require exposure to ionizing radiation. The technique therefore offers theopportunity to study large numbers of individuals and obtain measurements of surface facialstructures at a level of accuracy not possible with previous techniques. Such imaging hasbeen previously used to analyse craniofacial changes before and after various treatment20,21 modalities for OSA treatment, however to date only one study has used this technique toobtain 3D surface images of the face from 40 OSA and 40 non-OSA subjects, analysing only22 the association of craniofacial obesity with the OSA severity. No comprehensive study hasbeen undertaken with this technique to compare the discriminatory capacity of facialmorphometry between individuals with and without OSA. Most of the OSA prediction techniques use Linear Regression (LR) models based onbasic statistics. Machine learning-based classification algorithms are widely used in31 Biometrics , but have not been explored well for OSA feature classification. Recently, Sun33 et al. demonstrated that Genetic Algorithms performs better than LR model . However, theyonly used features extracted from clinical and polysomnography (PSG) measurements.SIGNIFICANCE AND INNOVATIONSignificanceClinically significant obstructive sleep apnoea (OSA) is a common condition occurring in 1- 1 2 2% of children and 2-4% of the middle-aged population. It is caused by repetitive upperairway obstruction due to collapse of upper airway structures during sleep. Obstructiverespiratory events are accompanied by repetitive oxygen desaturation and blood pressuresurges, and are usually terminated by brief awakenings (electroencephalographic arousals).3OSA impairs daytime function and is associated with major reductions in quality of life, 4 increased risk of motor vehicle accidents and cardiovascular disease; including hypertension,5 cardiac failure and stroke. According to a recent report of Deloitte Access Economics, in2010 the total financial and non-financial burden of OSA in Australia was estimated to be$21.2 billion, including direct health care cost of $575.42 million and indirect health care cost(due to lost productivity, deadweight loss, workplace/motor vehicle accidents, social security6 payments etc.) of $2.6 billion.The gold-standard assessment of OSA is laboratory-based PSG which is expensive(around $1800), time-consuming, requires sophisticated specialist facilities, technical and7scientific staff and sleep clinicians, which are commonly not available in all regions. Home- based sleep studies with rented devices costs less but has not been proved to be as accurate as8 9PSG. Due to the non-specific nature of symptoms associated with OSA and the limitedaccess to PSG, many OSA patients remain undiagnosed until significant symptom appears. It10is estimated that 82% of men and 92% of women with OSA have not been diagnosed. Byaddressing the diagnosis of this sleep disorder in a simpler and cheaper way using stateoftheart imaging and classification technologies, the project will cover one of the ninenational research priorities of Australia - Health.Innovation Most of the existing approaches use 2D features for OSA classifications/prediction. To thebest of our knowledge this project will be the first to propose comprehensive extraction of 3Dsurface features highly correlated to OSA severity.The use of automatic algorithms forfeature extraction and classification is innovative in this area of research. Combining threedifferent types (clinical, 2D image and 3D surface) of features is another avenue ofinnovation. APPROACH AND METHODOLOGYAim-1.Develop algorithms to extract automatically 2D and 3D craniofacial features thatpotentially phenotype OSA:Task-1. Raw Data Collection: There will be three datasets in this project: training,validation and testing. The machine learning-based classification algorithms will be trainedand validated using the training and validation sets respectively. The performance of thedeveloped algorithms will be evaluated on the testing set constituting samples not used in thetraining phase. The number of OSA subjects in these datasets will be 50, 25 and 50respectively and that of non-OSA (control) subjects will be 50, 25 and 25. OSA subjects forthis study will include individuals with a range of severities of OSA and already treated in theOral Health Centre of Western Australia (OHCWA), University of Western Australia'sCentre for Sleep Science (CSS)and ENT and Maxillofacial Surgery clinics of Fiona StanelyHospital and Hollywood Private Hospital. OHCWA and CSS has 3D scanner and 3Dphotographs are routinely taken from all clients/patients along with recording their clinicalobservation data. A portable eye-safe 3D scanner (to be purchased) will be used to collect 3Dphotographs of additional OSA patients from other clinics if required. Non-OSA subjects will be recruited from the students and staff of ECU. Afterobtaining ethics approval advertisement will be made for the recruitment. Interested volunteers will be first screened for likelihood of OSA via the Berlin and the EpworthSleepiness questionnaires and, an oral examination to measure pharyngeal grade andMallampatti score (assessments of pharyngeal crowding). Only those classified at low risk ofsleep apnoea with negligible daytime sleepiness and minimal pharyngeal crowding will berequested to undergo home sleep test (HST). They will be trained on how to use the HSTdevice. Prior to the sleep test, their 3D photograph of the face and neck areas will be capturedusing the portable 3D scanner. Task-2. Automatic Extraction of Facial Surface Features: The captured 3D photographswill be represented as a 3D surface (a triangulated polygon) mesh on a personal standarddesktop using MATLAB. Data will be rendered in a photo realistic model (3D textured) forvisual check. The face area will be detected automatically from the background using a veryfast and accurate face detection algorithm developed by Viola and Jones and used by CI in32 his biometric research. An extended window including head and neck will then be croppedfrom corresponding 3D surface data. Any surface defects such as ‘spikes’ or ‘holes’ will beautomatically refined using normalization algorithms developed by CI for ear and face32 biometrics. Following normalization, quantitative facial shape featureswill be extracted by CI and RA from the surface data. Tentativefeatures include length of the maxilla, mandible and chin, the(a) (b) Figure 2. (a) Facial surfacecircumference of the neck, and the relative shape ratios (RSRs) ofimages (b) Colour map ofthe superimposition of asome surface features (e.g. length of maxilla with respect to theface and its mirror showingfacial asymmetry. mandible and that of maxilla and mandible compared to the forehead38 and neck) proposed by CI and his collaborators. In order to extract these features34 automatically, Cascaded AdaBoost -based detection algorithms will be developed with "