Automatic Feature Extraction & Assessment Answer

IAH167

Literature Review Assignment Help:

Task:

Aims and Objectives
The diagnosis of Obstructive Sleep Apnoea (OSA), the most prevalent sleeping disorder, presents a significant problem to health care systems worldwide. Historically OSA diagnosis has been made with laboratory-based polysomnography (PSG) studies which are very time- consuming and expensive. Difficulties related to accessing PSG studies and the large number of individuals in the community who are thought to be undiagnosed has led to a substantial international effort to find alternative, simpler methods to detect OSA. The overriding aim of this project is to develop algorithms for a quick, safe and cost-effective way of predicting OSA for the early prevention of the syndrome. Computer Vision-based algorithms will be developed 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. These features along with other 2D and 3D image features recently demonstrated to be effective in the area of object detection (such as local descriptors and histogram of gradients) will then be passed to different a state of the art machine learning algorithms (including deep learning) for training. Exhaustive training on a large number of normal and sleep apnoeic samples would enable us to select the most distinctive features to be used for the prediction of moderate to severe levels of OSA. Specific aims of the project are to use surface and volumetric images of the head and neck to:

Aim-1 Develop algorithms to extract automatically 2D and 3D craniofacial features that potentially phenotype OSA.

Aim-2 Develop feature selection and classification algorithms to predict OSA.

Background
Many attempts have been made in the past to predict OSA based on questionnaires. For example, the Berlin questionnaire predicts the level of risk based on snoring, tiredness, blood pressure and body mass index information while the Epworth Sleepiness questionnaire assesses the sleepiness in various situations during the day. Although they are self-administered and low-cost, they have shortcomings in accurately identifying affected individuals. 9 A systematic review of eight different questionnaire models shows substantial variation in the diagnostic performance among them, none presenting reasonable sensitivity and specificity. 11

Imaging techniques have been considered as useful adjunctive tools to diagnose OSA, with the radiographic head film (cephalometric) analysis being the most convenient and widely used. 12 This technique has been used determining significant correlation between OSA severity and neck circumference in men, 13 retropalatal airway, 13 Upper Airway Length (UAL), 14,15 inferiorly placed hyoid bone, 15,16 facial depth, 17 mandibular plane angle 17 and reduced mid-face length. 16 Cone Beam Computed Tomography (CBCT), Medical CT, Magnetic Resonance Imaging are three dimensional counterpart of cephalometric images and hence provide more appealing results but they are more invasive and expensive. Recently, Lee et al. 19 have explored that digital photographs of craniofacial surface structures may be used to predict OSA. They analysed frontal and profile photographs of 114 subjects. Using only four photographic measurements (face width, eye width, cervicomental angle and mandibular length) their logistic regression based model obtained an accuracy of 76.1% with an area under Receiver Operating Characteristic (ROC) curve of 0.82. 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 or neck circumference) collected by clinical observations. However, like cephalometry, digital photographs are two dimensional in nature and hence neither non-linear measurements nor measurements of the shape of craniofacial anatomy can be obtained.
Three dimensional surface imaging technologies have recently been developed that are well suited for imaging the human head, face and neck. Images can be accurately obtained extremely quickly (less than one second) and the technique is non-invasive in nature as it does not require exposure to ionizing radiation. The technique therefore offers the opportunity to study large numbers of individuals and obtain measurements of surface facial structures at a level of accuracy not possible with previous techniques. Such imaging has been previously used to analyse craniofacial changes before and after various treatment modalities for OSA treatment, 20,21 however to date only one study has used this technique to obtain 3D surface images of the face from 40 OSA and 40 non-OSA subjects, analysing only the association of craniofacial obesity with the OSA severity. 22 No comprehensive study has been undertaken with this technique to compare the discriminatory capacity of facial morphometry between individuals with and without OSA.
Most of the OSA prediction techniques use Linear Regression (LR) models based on basic statistics. Machine learning-based classification algorithms are widely used in Biometrics 31 , but have not been explored well for OSA feature classification. Recently, Sun et al. demonstrated that Genetic Algorithms performs better than LR model 33 . However, they only used features extracted from clinical and polysomnography (PSG) measurements.

Aim-1. Develop algorithms to extract automatically 2D and 3D craniofacial features that potentially 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 trained and validated using the training and validation sets respectively. The performance of the developed algorithms will be evaluated on the testing set constituting samples not used in the training phase. The number of OSA subjects in these datasets will be 50, 25 and 50 respectively and that of non-OSA (control) subjects will be 50, 25 and 25. OSA subjects for this study will include individuals with a range of severities of OSA and already treated in the Oral Health Centre of Western Australia (OHCWA), University of Western Australia’s Centre for Sleep Science (CSS) and ENT and Maxillofacial Surgery clinics of Fiona Stanely Hospital and Hollywood Private Hospital. OHCWA and CSS has 3D scanner and 3D photographs are routinely taken from all clients/patients along with recording their clinical observation data. A portable eye-safe 3D scanner (to be purchased) will be used to collect 3D photographs of additional OSA patients from other clinics if required.

Non-OSA subjects will be recruited from the students and staff of ECU. After obtaining ethics approval advertisement will be made for the recruitment. Interested volunteers will be first screened for likelihood of OSA via the Berlin and the Epworth Sleepiness questionnaires and, an oral examination to measure pharyngeal grade and Mallampatti score (assessments of pharyngeal crowding). Only those classified at low risk of sleep apnoea with negligible daytime sleepiness and minimal pharyngeal crowding will be requested to undergo home sleep test (HST). They will be trained on how to use the HST device. Prior to the sleep test, their 3D photograph of the face and neck areas will be captured using the portable 3D scanner.

Task-2. Automatic Extraction of Facial Surface Features: The captured 3D photographs will be represented as a 3D surface (a triangulated polygon) mesh on a personal standard desktop using MATLAB. Data will be rendered in a photo realistic model (3D textured) for visual check. The face area will be detected automatically from the background using a very fast and accurate face detection algorithm developed by Viola and Jones and used by CI in his biometric research. 32 An extended window including head and neck will then be cropped from corresponding 3D surface data. Any surface defects such as ‘spikes’ or ‘holes’ will be automatically refined using normalization algorithms developed by CI for ear and face biometrics. 32

Following normalization, quantitative facial shape features will be extracted by CI and RA from the surface data. Tentative features include length of the maxilla, mandible and chin, the circumference of the neck, and the relative shape ratios (RSRs) of some surface features (e.g. length of maxilla with respect to the mandible and that of maxilla and mandible compared to the forehead and neck) proposed by CI and his collaborators. 38 In order to extract
these features automatically, Cascaded AdaBoost 34 -based detection algorithms will be developed with exhaustive training with positive and negative samples of the anatomic components. The approach will be similar to that CI developed for ear detection. 32 Some surface features will also be extracted based on the analysis of the surface difference (see Fig 2(b)) Figure 2. (a) Facial surface images (b) Colour map of the superimposition of a face and its mirror showing facial asymmetry.

between average faces 20 of OSA and non-OSA subjects of different ages and genders Average faces will be constructed following an approach similar to that of my previous collaborator Prof Clement’s group 35 which was shown reliable even if constructed from a small number of samples.

Aim-2. Develop feature selection and classification algorithms to predict OSA:

Task-3. Feature Selection: Extracted features will be reduced to a minimum number by only choosing those features that have most discriminating attributes. In this process, we will adopt a cluster-classification technique36 used to reduce the extraneous features from the outer layers of each cluster by constantly monitoring the noteworthy attributes. It was shown to be efficient in identifying the important features in a large dataset to be selected as representatives of the entire dataset without significant loss of information. At first, appropriate noise filtering and condensation algorithms will be developed for prototype selection. Then an efficient prototype construction method will be used to find the new attributes that can represent the whole data ore compactly.

Task-4. Feature Classification and OSA Prediction: A multi-level classification technique will be developed using Machine Learning (ML) algorithms including to classify the extracted features. In this method each training sample belongs to one of the two different classes (positive or negative) and the goal is to construct a function which, given a new sample, will correctly predict the class to which it belongs.

We will adopt a variant of k-fold cross validation (k-FCV) that ensures each of the input vectors is tested. 37 At first one group of data from a particular data set (S1) will be selected and divided into k portions of equal size; then another group of data (S2) of the same size as (S1) will be selected and also divided it into k portions, as indicated in Fig. 3. To establish the training set, proposed detection engine will take k-1 portions from each of the data groups (S1
& S2) and the remaining portions from both groups (S1 & S2) are used for the testing set. As is customary, the training set will be used to establish the model and the testing set will be used to validate it. The whole process will be repeated so that every portion of both groups (S1 & S2) is chosen as testing data; the results will then be averaged in order to ensure that the input vectors are trained and tested over a broad spectrum of classifiers. Different machine learning algorithms including a boosting technique (AdaBoost) and deep learning will be used and their performance will be compared.

Task-5. Performance Evaluation: The performance of the developed algorithms will be evaluated on the testing set comprising 25 OSA and 25 non-OSA subjects. Prediction accuracy curves (including receiver operating characteristic (ROC)) will be generated to show up the results. Accuracy of using different types of features separately and in combination will be tested in order to see whether surface features alone or in combined withselected clinical observatory features can predict OSA with sufficient accuracy without requiring volumetric imaging.

Task-6. Software Development: Algorithms will be developed and tested using MATLAB. A Graphical User Interface will be developed using Developer with Oracle database support which will facilitate entering and storing patient’s demographic/clinical observatory data and imageries for further processing.

IMPACT AND NATIONAL BENEFIT
Improving Quality of Life: OSA patients are not only deprived from sound sleep, they affect sleep partners and through their daytime sleepiness affect the whole society around them. The proposed conservative, cost-effective, patient-convenient and widely applicable methods to predict OSA will facilitate more accessible diagnosis of a larger number of patient groups than is possible with polysomnography and will enhance early intervention. Thus this
work has the potential to have an enormous impact on the quality of lives of thousands of Australians.

Reducing Healthcare Cost: As mentioned earlier, an undiagnosed OSA patients incur additional healthcare cost associated with their poor health condition which is estimated to US $1,336 per annum. 39 Facilitating early intervention of OSA through the proposed affordable methods will help reducing this cost. Facilitating Training: This research would allow me to co-ordinate and perform research in a vibrant, multidisciplinary area and develop myself into an independent researcher in the field of 3D Imaging. The facilitated collaboration between inter-school and inter-university experts would increase in the training of postgraduate students.

COMMUNICATION OF RESULTS
The AIORC group of ECU School of Science, UWA VVR group at CSSE and UWA Orthodontics Research group have made significant research contributions over a number of years through patents, publications, PhD dissertations, as well as considerable media attentions. All the Schools always encourage conference attendance and presentation of
research at different forums and exhibitions, school’s weekly seminars and yearly conferences. They also encourage collaborating and building networks with researchers from home and abroad.