Browsing by Author "Mutsvangwa, Tinashe"

Now showing 1 - 17 of 17
  • Thumbnail Image
    Item
    Open Access
    A deep learning algorithm for contour detection in synthetic 2D biplanar X-ray images of the scapula: towards improved 3D reconstruction of the scapula
    (University of Cape Town, 2020) Namayega, Catherine; Mutsvangwa, Tinashe; Malila, Bessie; Douglas, Tania
    Three-dimensional (3D) reconstruction from X-ray images using statistical shape models (SSM) provides a cost-effective way of increasing the diagnostic utility of two-dimensional (2D) X-ray images, especially in low-resource settings. The landmark-constrained model fitting approach is one way to obtain patient-specific models from a statistical model. This approach requires an accurate selection of corresponding features, usually landmarks, from the bi-planar X-ray images. However, X-ray images are 2D representations of 3D anatomy with super-positioned structures, which confounds this approach. The literature shows that detection and use of contours to locate corresponding landmarks within biplanar X-ray images can address this limitation. The aim of this research project was to train and validate a deep learning algorithm for detection the contour of a scapula in synthetic 2D bi-planar Xray images. Synthetic bi-planar X-ray images were obtained from scapula mesh samples with annotated landmarks generated from a validated SSM obtained from the Division of Biomedical Engineering, University of Cape Town. This was followed by the training of two convolutional neural network models as the first objective of the project; the first model was trained to predict the lateral (LAT) scapula image given the anterior-posterior (AP) image. The second model was trained to predict the AP image given the LAT image. The trained models had an average Dice coefficient value of 0.926 and 0.964 for the predicted LAT and AP images, respectively. However, the trained models did not generalise to the segmented real X-ray images of the scapula. The second objective was to perform landmark-constrained model fitting using the corresponding landmarks embedded in the predicted images. To achieve this objective, the 2D landmark locations were transformed into 3D coordinates using the direct linear transformation. The 3D point localization yielded average errors of (0.35, 0.64, 0.72) mm in the X, Y and Z directions, respectively, and a combined coordinate error of 1.16 mm. The reconstructed landmarks were used to reconstruct meshes that had average surface-to-surface distances of 3.22 mm and 1.72 mm for 3 and 6 landmarks, respectively. The third objective was to reconstruct the scapula mesh using matching points on the scapula contour in the bi-planar images. The average surface-to-surface distances of the reconstructed meshes with 8 matching contour points and 6 corresponding landmarks of the same meshes were 1.40 and 1.91 mm, respectively. In summary, the deep learning models were able to learn the mapping between the bi-planar images of the scapula. Increasing the number of corresponding landmarks from the bi-planar images resulted into better 3D reconstructions. However, obtaining these corresponding landmarks was non-trivial, necessitating the use of matching points selected from the scapulae contours. The results from the latter approach signal a need to explore contour matching methods to obtain more corresponding points in order to improve the scapula 3D reconstruction using landmark-constrained model fitting.
  • Thumbnail Image
    Item
    Open Access
    A structured light solution for detecting scapular dyskinesis
    (2018) Verster, Jaco; Gray, Janine; Sivarasu, Sudesh; Mutsvangwa, Tinashe
    Scapular dyskinesis is a common occurrence in overhead athletes, i.e. athletes who participate in any sport where the upper arm and shoulder is used above the athlete’s head. However, no consensus has been reached on how to evaluate scapular dyskinesis quantitatively. In this thesis, we developed a measuring tool that can be used to evaluate certain key clinical parameters specific to scapular dyskinesis. The tool employs a 3D structured light computer vision approach to create a surface map of the soft-tissue across the scapula. This surface map is then analysed using surface curvature analysis techniques to identify the key clinical parameters associated with scapular dyskinesis. The main advantage of this method is that it provides a measurement tool that may facilitate future quantitative analysis of these key parameters. This may aid with diagnosis and monitoring of the condition by allowing measurement data to be collected both before and after treatment and rehabilitation. We expect that this tool will make the monitoring of treatment effectiveness easier while contributing to diagnostic computer vision.
  • Thumbnail Image
    Item
    Open Access
    Articulated Statistical Shape Modelling of the Shoulder Joint
    (2020) Alemneh, Tewodros; Mutsvangwa, Tinashe; Douglas, Tania
    The shoulder joint is the most mobile and unstable joint in the human body. This makes it vulnerable to soft tissue pathologies and dislocation. Insight into the kinematics of the joint may enable improved diagnosis and treatment of different shoulder pathologies. Shoulder joint kinematics can be influenced by the articular geometry of the joint. The aim of this project was to develop an analysis framework for shoulder joint kinematics via the use of articulated statistical shape models (ASSMs). Articulated statistical shape models extend conventional statistical shape models by combining the shape variability of anatomical objects collected from different subjects (statistical shape models), with the physical variation of pose between the same objects (articulation). The developed pipeline involved manual annotation of anatomical landmarks selected on 3D surface meshes of scapulae and humeri and establishing dense surface correspondence across these data through a registration process. The registration was performed using a Gaussian process morphable model fitting approach. In order to register two objects separately, while keeping their shape and kinematics relationship intact, one of the objects (scapula) was fixed leaving the other (humerus) to be mobile. All the pairs of registered humeri and scapulae were brought back to their native imaged position using the inverse of the associated registration transformation. The glenohumeral rotational center and local anatomic coordinate system of the humeri and scapulae were determined using the definitions suggested by the International Society of Biomechanics. Three motions (flexion, abduction, and internal rotation) were generated using Euler angle sequences. The ASSM of the model was built using principal component analysis and validated. The validation results show that the model adequately estimated the shape and pose encoded in the training data. Developing ASSM of the shoulder joint helps to define the statistical shape and pose parameters of the gleno humeral articulating surfaces. An ASSM of the shoulder joint has potential applications in the analysis and investigation of population-wide joint posture variation and kinematics. Such analyses may include determining and quantifying abnormal articulation of the joint based on the range of motion; understanding of detailed glenohumeral joint function and internal joint measurement; and diagnosis of shoulder pathologies. Future work will involve developing a protocol for encoding the shoulder ASSM with real, rather than handcrafted, pose variation.
  • No Thumbnail Available
    Item
    Open Access
    Automated analysis of digital medical images in cervical cancer screening: A systematic review
    (2023) Ledwaba, Leshego; Mutsvangwa, Tinashe
    Background Cervical cancer is the second highest cause of mortalities in women living in resource-constrained countries compared to those living in high income countries, due to lack of organized population screening. Cervical cancer screening is the best way to detect lesions and remove them before they advance into malignancy. In South Africa, the current standard cervical cancer screening protocol begins with cytology examination then a referral for colposcopy follows if the cytology screening test is abnormal. Thereafter, histopathology examination is conducted on biopsy specimen collected during colposcopy. Biopsy specimen are only collected when suspicious lesions are observed during colposcopy. These screening procedures need to be performed by qualified specialist clinicians because accuracy of diagnosis is highly dependent on the skill level and experience of the clinician making the diagnostic decision. In South Africa, colposcopy and cytology are constrained by a shortage of specialists and expensive diagnostic infrastructure. Consequently, public health interventions such as population screening programs for cervical cancer are poorly implemented. Researchers have been developing low-cost portable devices, some of which are incorporated with automated image analysis to enhance diagnostic decision-making. The methods for incorporating automation within each domain of the cervical cancer screening protocol are becoming numerous as researchers continue to advance the field. As the knowledge base is growing rapidly, progress on the implementation status of novel imaging devices and novel algorithms in cervical cancer screening has become unclear. Thus, there is a need to identify all relevant technologies, i.e. devices and algorithms, currently being researched in the field, and to understand their unique strengths and challenges toward clinical adoption. The aim of this project was to provide a systematic review summarizing the full range of automated technology systems used in cervical cancer screening. Method A systematic search on five main academic databases (PubMed, Scopus, EBSCOhost, Web of Science and Google Scholar) was conducted to identify articles on automated technology systems applied in cervical cancer screening. The search results were screened by two independent reviewers to assess eligibility in meeting Population, Intervention, Comparator, and Outcome (PICO) criteria. The screening of articles was a two-step approach: firstly, screening for eligibility by only reading the title and abstract of articles; then secondly, screening by reading full texts. A data extraction form was developed and used to systematically summarize information contained in 70 studies that were included for analysis. Bias in each study was assessed using a risk of bias template adapted from established checklists, namely the Cincinnati Children's LEGEND guideline and the Joanna Briggs Institute critical appraisal checklist for diagnostic test accuracy studies. A conceptual map of common computer aided diagnostics (CAD) tasks that make up the automation pipeline was developed as a narrative tool to synthesize the specific functions that proposed CAD algorithms in multiple screening domains were performing. Results This systematic review found 16 studies which reported application of algorithms paired with novel image acquisition devices, and 52 studies reporting on standalone image analysis algorithms. CAD algorithms associated with acquisition devices (both novel and conventional) revealed that automated analysis achieved superior performance than manual expert analysis; thus, improving diagnostic decisions made by clinicians performing colposcopy, cytology and histopathology. The pertinent algorithms were those developed for devices designed with a mobile phone or tablet, which were the Pocket Colposcope, MobileODT EVA Colpo, Smartphone Camera, Smartphone-based Endoscope System, Smartscope, Mobile high resolution micro-endoscopy (mHRME), and Pi high resolution micro-endoscopy (PiHRME). These mobile-based systems in particular could be applied more widely in low- to middle-income countries than bulky devices because of their anticipated low purchase cost. Most interventions were in the feasibility stage of development, undergoing initial clinical validations. Conclusion This review found that cervical cancer screening researchers have proven the positive clinical impact that CAD algorithms might have in reaching outstanding prediction performance. This accomplishment is a significant step toward minimizing reliance on experts to provide cervical cancer screening services. Furthermore, the systematic review summarized evidence of the algorithms which are being created utilizing portable devices, to circumvent constraints prohibiting wider implementation in LMICs (such as expensive diagnostic infrastructure). These advances can make the decentralization of colposcopy services more feasible if unsupervised community health workers are trained to effectively utilize portable imaging devices with automated functionality for interpreting results. However clinical validation of promising novel systems is not yet implemented adequately in LMICs, because most studies did not include nurses who are a crucial segment of the target population. Additionally, it is not clear whether the proposed portable interventions are economically feasible when hidden costs are also taken into account.
  • Thumbnail Image
    Item
    Open Access
    Development of a statistical shape and appearance model of the skull from a South African population
    (2018) Lugadilu, Brian; Mutsvangwa, Tinashe; Douglas, Tania
    Statistical shape models (SSMs) and statistical appearance models (SAMs) have been applied in medical analysis such as in surgical planning, finite element analysis, model-based segmentation, and in the fields of anthropometry and forensics. Similar applications can make use of SSMs and SAMs of the skull. A combination of the SSM and SAM of the skull can also be used in model-based segmentation. This document presents the development of a SSM and a SAM of the human skull from a South African population, using the Scalismo software package. The SSM development pipeline was composed of three steps: 1) Image data segmentation and processing; 2) Development of a free-form deformation (FFD) model for establishing correspondence across the training dataset; and 3) Development and validation of a SSM from the corresponding dataset. The SSM was validated using the leave one-out cross-validation method. The first eight principal components of the SSM represented 92.13% of the variation in the model. The generality of the model in terms of the Hausdorff distance between a new shape generated by the SSM and instances of the SSM had a steady state value of 1.48mm. The specificity of the model (in terms of Hausdorff distance) had a steady state value of 2.04mm. The SAM development pipeline involved four steps: 1) Volumetric mesh generation of the reference mesh to be used in establishing volumetric correspondence; 2) Sampling of intensity values from original computed tomography (CT) images using the in-correspondence volumetric meshes; and 3) Development of a SAM from the in-correspondence intensity values. A complete validation of the SAM was not possible due to limitations of the Scalismo software. As a result, only the shapes of the incomplete skulls were reconstructed and thereby validated. The amount of missing detail, as represented by absent landmarks, affected the registration results. Complete validation of the SAM is recommended as future work, via the use of a combined shape and intensity model (SSIM).
  • Thumbnail Image
    Item
    Restricted
    Development of an assessment framework for student engagement in design thinking projects for health innovation
    (2020) Dikgomo, Kagiso; Douglas, Tania S; Mutsvangwa, Tinashe; Hendricks, Sharief
    Student engagement is a dynamic and multifaceted concept – it encompasses physical, emotional, and cognitive components. Various instruments to assess student engagement exist, however these are not intended to assess how students engage with one another and with community stakeholders in participatory health projects. Although instruments do exist to assess participation/power-sharing in participatory health projects, none of the available instruments are suitable for the assessment of student engagement in such projects. The current study set out to develop an assessment framework for student engagement in design thinking projects for health innovation. Design thinking is a human-centred and participatory approach to problem-solving. The objectives of the project were: (1) the design and implementation of a questionnaire to assess student engagement in design thinking activities, and (2) assessment of the validity of the questionnaire. A preliminary questionnaire was developed with the aid of the literature and implemented for students taking a postgraduate course called Health Innovation & Design, which follows a design thinking approach for health innovation. Analysis of students’ written reflective reports and a focus group discussion were used to revise the questionnaire items. The revised questionnaire was validated by design thinking practitioners (the course lecturer and facilitators), and further modifications were made based on their views. The assessment framework developed in this study incorporates the design thinking phases according to the IDEO design thinking approach, an educational definition of student engagement, and recommendations by students of the Health Innovation & Design course and their course lecturer and facilitators. This questionnaire may be used to assess engagement in academic settings as well as non-academic settings when design thinking is applied for health innovation.
  • No Thumbnail Available
    Item
    Open Access
    Evaluating the influence of machine-specific DRR parameters on the accuracy of X-ray simulation and orthopaedic 2D-3D reconstruction
    (2023) Reyneke, Cornelius; Mutsvangwa, Tinashe; Douglas Tania
    In orthopaedics, two-dimensional-to-three-dimensional (2D-3D) reconstruction allows 3D bone structures, conventionally derived from 3D modalities such as computed tomography (CT), to be derived from 2D modalities such as X-ray imaging. Thus, clinical interventions such as implant design and postoperative evaluation can be made more accessible, less expensive and, in some cases, the dose of ionising radiation to the patient can be reduced. State-of-the-art approaches iteratively warp a deformable 3D model of the bone, simulate an X-ray projection image from it, and compare the result to a real X-ray target image, with the goal of minimising the disparity between the two. The X-ray simulation method includes implicit X-ray machine-specific calibration settings, which affect both the resulting geometry and intensity profile of simulated X-ray images. However, the importance of correct projection calibration is not adequately discussed in literature, despite the fact that X-ray machines vary significantly with regard to their imaging setup. Furthermore, the reconstruction inaccuracies resulting from projection miscalibration are not well understood. In this thesis, first, an extensive literature review of the 2D/3D reconstruction problem is conducted and a unified mathematical formulation is proposed. Next, the development of a digitally reconstructed radiograph (DRR) renderer for simulating X-ray images is reported. The DRR renderer was developed using a standard volume rendering framework, adapted to the unique requirements of X-ray imaging. The renderer can be calibrated to machine-specific parameters, of which two were the focus in-depth experimental investigations: the distance from the X-ray source to the imaged object (S2O distance), and modelled energy intensity of the X-ray source (incident energy). For each correctly calibrated parameter of a nominal value, a set of perturbed values were generated and used to render corresponding simulated X-ray images. The resulting rendering errors were then measured through comparison to a groundtruth X-ray image of an object with known geometry and density. The results demonstrated a 2D Dice overlap error of 0.4% for every 1 cm with which the S2O distance was offset, and a pixel intensity error of 1% for every 0.1 point offset of the incident energy coefficient. In a subsequent experimental investigation, the same two machine-specific parameters were studied to determine the influence of machine-specific parameters on 2D-3D reconstruction accuracy using a single image. First, having generated a set of random DRRs, each was viii used as a simulated registration target while offsetting the correctly calibrated parameters in the same fashion as for the first experiment. A 2D Dice overlap error of at least 0.3% was estimated for every 1 cm with which the S2O distance was offset. Similar experiments were performed on real X-ray images of dry-bone femurs as registration targets, leading to a 2D Dice overlap error of at least 0.4% for every 1 cm with which the S2O distance was offset. The incident energy parameter had no discernible impact on the resulting registration accuracy for either the simulated or real-world targets. As a whole, the experimental results demonstrate that machine-specific calibration has a noticeable impact on the accuracy of simulated X-ray images and, in the case of the S2O distance, also on the accuracy of 2D-3D reconstruction. An in-depth commentary on the work done, together with suggestions for future research, concludes the thesis. In summary, the presented work provides valuable insight into previously overlooked facets of orthopaedic 2D-3D reconstruction, and suggests that machine-specific calibration should be carefully considered when performing 2D-3D reconstruction.
  • Thumbnail Image
    Item
    Open Access
    Feature detection in ultrasound images for computer aided diagnosis of Hodgkin's Lymphoma
    (2021) Dawood, Tareen; Mutsvangwa, Tinashe; Verburgh, Estelle
    The varying clinical presentation of Hodgkin's lymphoma (HL) poses a diagnostic challenge in South Africa, as the clinical picture of this lymphoma overlaps with prevalent comorbidities such as tuberculosis (TB) and the Human Immuno-Deficiency Virus (HIV). HIV infection additionally increases the risk of developing HL. These factors motivate for the need to investigate the role of imaging modalities in the diagnostic pathway of HL. The goal of this project was to develop and evaluate an automated framework for improving diagnostic imaging interpretability of ultrasound for HL diagnosis in a HIV TB endemic environment. To achieve this, a precise abdominal ultrasound protocol was developed with clinical guidance. The specific frames in the protocol were used to detect several image biomarkers of clinical interest: splenic enlargement (splenomegaly), splenic lesions, splenic microabscesses, abdominal lymph node enlargement, ascites, and effusions (pleural and pericardial). The developed protocol provided a novel guideline to identify an abnormality from the available ultrasound images. A secondary outcome of the protocol was the development of a prospective guide to image Hodgkin's lymphoma patients using ultrasound, however further testing and evaluation is required to validate its use. Image processing techniques were then applied to identified frames, and geometrical and textural features extracted, to develop an automated abnormality characterisation framework. A total of 36 features were extracted and used to characterise each abnormality. Thereafter, an automated algorithm was used to characterise and classify Hodgkin's lymphoma. A support vector machine model was built, with two experiments performed to evaluate the model. The model achieved a maximum training accuracy of 83%, similar in performance to support vector machine classification models used in medical applications. Noticeably the classification accuracy increased favourably when specific abnormalities were assessed: an enlarged spleen, splenic micro abscesses, ascites, pleural effusions, and pericardial effusions. This may indicate that these specific abnormalities are sufficient to differentiate patients with and without Hodgkin's lymphoma but understanding the reasoning for the decision taken by the system requires further investigation. In this study we show how image processing and automated classification techniques when applied to ultrasound images, have the potential to improve the differential diagnostic pathway of HL. Further evaluation using a larger dataset is planned, to validate and implement these findings in a strained healthcare setting.
  • Thumbnail Image
    Item
    Open Access
    Geometric morphometrics for 3D dense surface correspondence: population comparisons of shoulder bone morphology
    (2018) Fouefack, Jean-Rassaire; Mutsvangwa, Tinashe; Douglas, Tania; Inyang, Adijat Omowumi
    Background: Comparisons in morphological shape/form across population groups could provide population differences that might assist in making decisions on diagnosis and prognosis by the clinician. Geometric morphometrics (GM) is one of the fields that help to provide such population comparisons. In medical imaging and related disciplines, GM is commonly done using annotated landmarks or distances measured from 3D surfaces (consisting of triangular meshes). However, these landmarks may not be sufficient to describe the complete shape. This project aimed to develop GM for analysis that consider all vertices in the triangular mesh as landmarks. The developed methods were applied to South African and Swiss shoulder bones (scapula and humerus) to analyse morphological differences. Methods: The developed pipeline required first establishing correspondence across the datasets through a registration process. Gaussian process fitting was chosen to perform the registration since it is considered state-of-the-art. Secondly, a novel method for automatic identification of vertices or areas encoding the most shape/form variation was developed. Thirdly, a principal component analysis (PCA) that addressed the high dimensionality and lower sample size (HDLSS) phenomenon was adopted and applied to the dense correspondence data. This approach allowed for the stabilisation of the distribution of the data in low-dimensional form/shape space. Lastly, appropriate statistical tests were developed for population comparisons of the shoulder bones when dealing with HDLSS data in both form and shape space. Results: When the mesh-based GM analysis approach was applied to the training datasets (South African and Swiss shoulder bones), it was found that the anterior glenoid which is often the site of the shoulder dislocation is the most varied area of the glenoid. This has implications for diagnosis and provides knowledge for prosthesis design. The distribution of the data in the modified PCA space was shown to converge to a stable distribution when more vertices/landmarks are used for the analysis. South African and Swiss datasets were shown to be more distinguishable in a low-dimensional space when considering form rather than shape. It was found that left and right South African scapula bones are significantly different in terms of shape. Discussion: In general, it was observed that the two populations means can be significantly different in shape but not in form. An improved understanding of these observed shape and form differences has utility for shoulder arthroplasty prosthesis design and may also be useful for orthopaedic surgeons during surgical preoperative planning.
  • Thumbnail Image
    Item
    Open Access
    Image analysis for a mobile phone-based assessment of latent tuberculosis infection
    (University of Cape Town, 2020) Maclean, Sarah; Mutsvangwa, Tinashe; Malila, Bessie; Douglas, Tania
    The current, most widely used method to screen for latent tuberculosis infection is the Mantoux tuberculin skin test, where tuberculin is injected into a patient's arm and may result in a cutaneous induration forming at the site of injection. A diameter measurement of the resultant induration, recorded using a ruler and ball point pen, is currently used to indicate the presence of latent tuberculosis infection. Limitations associated with the tuberculin skin test procedure are the crudeness of the induration measurement method, the follow-up clinical visit required from patients to have their induration measured, and the need for trained clinicians who can perform the induration measurement. These limitations motivated research into a mobile phone-based screening system which can be used to obtain a more accurate measurement of the induration without the need for a second visit to the clinic by patients. The prototype screening tool consists of a user interface for capturing induration images and a backend processing system that produces a threedimensional reconstruction of the induration for measurement. Recommendations from previous studies on the prototype screening tool, which involved evaluation of the mobile application using mock induration images, included improving the accuracy of measuring the induration and evaluating the tool on real induration images. The aim of this study was to evaluate the existing backend system and explore an alternative assessment approach for assessing the induration. This was achieved through the following objectives: (1) applying the current backend system to real induration images, (2) examining the need for three-dimensional reconstruction for delineation of the induration for measurement and (3) exploring an alternative method for the assessment of induration images using deep learning. Results for the first objective showed the three-dimensional reconstruction to be unsuccessful on real images. This was due to the homogeneity between the indurations and the surrounding skin, rendering the algorithm ineffective in delineating the indurations to obtain the diameter measurement required for diagnosis. The second objective involved determining whether the image orientation or induration height affected the diagnostic measurement. It was found that real indurations are much flatter and more subtle compared to the mock indurations used in the previous studies. This motivated an alternative image assessment approach using deep learning. However, deep learning approaches require large databases of annotated images to prevent overfitting on training data. The last objective therefore involved the design and implementation of a generative adversarial network for generation of synthetic images from a limited number of real images, which allowed the generation of an unlimited number of realistic-looking synthetic images from 150 real induration images.
  • Thumbnail Image
    Item
    Open Access
    Modelling non-linearity in 3D shapes: A comparative study of Gaussian process morphable models and variational autoencoders for 3D shape data
    (2021) Fehr, Fabio; Clark, Allan; Mutsvangwa, Tinashe
    The presence of non-linear shape variation in 3D data is known to influence the reliability of linear statistical shape models (SSM). This problem is regularly acknowledged, but disregarded, as it is assumed that linear models are able to adequately approximate such non-linearities. Model reliability is crucial for medical imaging and computer vision tasks; however, prior to modelling, the non-linearity in the data is not often considered. The study provides a framework to identify the presence of non-linearity in using principal component analysis (PCA) and autoencoders (AE) shape modelling methods. The data identified to have linear and non-linear shape variations is used to compare two sophisticated techniques: linear Gaussian process morphable models (GPMM) and non-linear variational autoencoders (VAE). Their model performance is measured using generalisation, specificity and computational efficiency in training. The research showed that, given limited computational power, GPMMs managed to achieve improved relative generalisation performance compared to VAEs, in the presence of non-linear shape variation by at least a factor of six. However, the non-linear VAEs, despite the simplistic training scheme, presented improved specificity generative performance of at least 18% for both datasets.
  • No Thumbnail Available
    Item
    Metadata only
    Multi-object and multi-feature models of thumb anatomy for population based morphological assessment
    (2024) Farrell, Caitlin; Mutsvangwa, Tinashe; Vereecke, Evie; Borotikar, Bhushan
    The trapeziometacarpal (TMC) joint, formed by the junction of the trapezium and first metacarpal (MC1) bones, is highly susceptible to the onset and progression of osteoarthritis (OA). Although various treatment options exist to ease symptoms and slow the progression of OA, the success of these treatments is heavily reliant on the early diagnosis of the disorder. The primary aim of this research was to develop a methodology for the use of computed tomography (CT) imaging in the characterisation of biomechanical risk factors of OA in the TMC joint using data from OA-affected and control subjects. Features related to shape, pose, and intensity in the CT images of the TMC joints from the subjects were correlated to a range of biomechanical risk factors. Multi-object and multi feature-class statistical models of control and OA-affected datasets were created through the development of 3 pipelines – a registration pipeline, a correspondence pipeline, and a model building pipeline. These models were used to compare the ranges of 5 morphological anatomical measures, namely: the distances between the articular surfaces of the bones; the angle of volar beak protrusion; the angle of trapezial inclination; the surface areas of the articular facets; and the radii of curvature. In line with the respective hypotheses, the distance between the articular surfaces and the angle of volar beak protrusion were seen to decrease in OA-affected data whilst the surface areas of the articular facets were seen to increase. Contrary to their respective hypotheses, the angle of trapezial inclination was seen to decrease in OA-affected data and the radii of curvature of the articular facets showed no significant changes in morphology. This suggests certain anatomical measures may be indicative of the onset and progression of TMC OA and provides a range of values typical of both control subjects and OA-affected subjects. A third model representative of the combined OA-affected and control data was developed and used to determine the correlations between the shape and pose, and the shape and intensity feature classes. This indicated a high correlation between both the shape and intensity and the shape and pose feature classes, thus suggesting a correlation between the respective biomechanical risk factors. In summary, the results of this research suggest that a decreased distance between the articular surfaces, an increased articular surface area, and a decrease in the angle of volar beak protrusion may be indicative of the onset and progression of TMC OA. This research is limited by the relatively small size of the datasets used and thus further research is necessary to determine the variation in the trapezial angle of inclination and the change in concavity of the articular surface of the MC1. Moreover, this research serves as an explanation and demonstration of the developed pipelines in the creation of a DMFC-GPM of the TMC joint which can be applied to larger and more diverse datasets in future research to determine the correlations between the respective feature classes.
  • Thumbnail Image
    Item
    Open Access
    Reconstruction of three-dimensional facial geometric features related to fetal alcohol syndrome using adult surrogates
    (2020) Atuhaire, Felix; Mutsvangwa, Tinashe; Douglas, Tania
    Fetal alcohol syndrome (FAS) is a condition caused by prenatal alcohol exposure. The diagnosis of FAS is based on the presence of central nervous system impairments, evidence of growth abnormalities and abnormal facial features. Direct anthropometry has traditionally been used to obtain facial data to assess the FAS facial features. Research efforts have focused on indirect anthropometry such as 3D surface imaging systems to collect facial data for facial analysis. However, 3D surface imaging systems are costly. As an alternative, approaches for 3D reconstruction from a single 2D image of the face using a 3D morphable model (3DMM) were explored in this research study. The research project was accomplished in several steps. 3D facial data were obtained from the publicly available BU-3DFE database, developed by the State University of New York. The 3D face scans in the training set were landmarked by different observers. The reliability and precision in selecting 3D landmarks were evaluated. The intraclass correlation coefficients for intra- and inter-observer reliability were greater than 0.95. The average intra-observer error was 0.26 mm and the average inter-observer error was 0.89 mm. A rigid registration was performed on the 3D face scans in the training set. Following rigid registration, a dense point-to-point correspondence across a set of aligned face scans was computed using the Gaussian process model fitting approach. A 3DMM of the face was constructed from the fully registered 3D face scans. The constructed 3DMM of the face was evaluated based on generalization, specificity, and compactness. The quantitative evaluations show that the constructed 3DMM achieves reliable results. 3D face reconstructions from single 2D images were estimated based on the 3DMM. The MetropolisHastings algorithm was used to fit the 3DMM features to 2D image features to generate the 3D face reconstruction. Finally, the geometric accuracy of the reconstructed 3D faces was evaluated based on ground-truth 3D face scans. The average root mean square error for the surface-to-surface comparisons between the reconstructed faces and the ground-truth face scans was 2.99 mm. In conclusion, a framework to estimate 3D face reconstructions from single 2D facial images was developed and the reconstruction errors were evaluated. The geometric accuracy of the 3D face reconstructions was comparable to that found in the literature. However, future work should consider minimizing reconstruction errors to acceptable clinical standards in order for the framework to be useful for 3D-from-2D reconstruction in general, and also for developing FAS applications. Finally, future work should consider estimating a 3D face using multi-view 2D images to increase the information available for 3D-from-2D reconstruction.
  • Thumbnail Image
    Item
    Open Access
    Three-Dimensional Body Volume Measurement From Two-Dimensional Images: Towards A Smartphone Application
    (2020) Majola, Khwezi; Mutsvangwa, Tinashe; Douglas, Tania; Lambert, Vicki
    Obesity poses a public health threat worldwide and is associated with a higher mortality, increased likelihood of diabetes, and an increased risk of cancer. When treating obesity, regular monitoring of metrics such as body mass index (BMI) and waist circumference has been found to result in improved health outcomes for patients. Three-dimensional (3D) scanners provide a useful tool to provide body measurements based on 3D images in obesity management. However, such scanners are often inaccessible due to cost. A smartphone image-based method able to produce 3D images may provide a more accessible measuring tool. As a step towards developing such a smartphone application, this project developed a method for 3D reconstruction of body images from two-dimensional (2D) images, using a full body 3D Gaussian process morphable model (GPMM). Separate GPMMs were trained to learn the shape of female and male human bodies. Gaussian process regression of the three-dimensional (3D) GPMM models onto two-dimensional (2D) images is performed. Corresponding landmarks on the 3D shapes and in the 2D images are employed in reconstruction. Measurements of body volume, waist circumference and height are then performed to extract information that is useful in obesity management. Different model configurations (shape model with arms; modified shape model with arms; shape model without arms; marginalised shape model without arms; shape model with different landmarks) were used to ascertain the most promising approach for the reconstruction. Each reconstructed body was tested for accuracy using the surface-tosurface distance per vertex, modified Hausdorff distance, and assessment of the measurements. Tests were performed using data from the same dataset used to build the model and generalised data from a different dataset. In all test cases, the best performing approach used shape models without arms when considering surface distances. However, the surface-to-surface distances errors were larger than those seen in literature. For body measurements, the best performing models varied with different models performing best for different measurements. For the measurements, the errors were larger than the allowable errors and larger than those found in literature. Landmark positions were evaluated separately and found to be imprecise. There are a few sources that contribute towards the reconstruction errors. Possible sources of error include an inability to interpret pose and landmark position errors. The major recommendations for future work are to use a model that incorporates both shape and pose and to use automatic landmarking methods. Regarding a pathway to a smartphone app, camera parameter information should be considered to improve processing of the images and smartphone orientation information should be considered to correct for distortions due to a tilted phone.
  • Thumbnail Image
    Item
    Open Access
    Towards a framework for multi class statistical modelling of shape, intensity and kinematics in medical images
    (2021) Fouefack, Jean-Rassaire; Mutsvangwa, Tinashe; Burdin, Valérie; Douglas, Tania; Borotikar, Bhushan
    Statistical modelling has become a ubiquitous tool for analysing of morphological variation of bone structures in medical images. For radiological images, the shape, relative pose between the bone structures and the intensity distribution are key features often modelled separately. A wide range of research has reported methods that incorporate these features as priors for machine learning purposes. Statistical shape, appearance (intensity profile in images) and pose models are popular priors to explain variability across a sample population of rigid structures. However, a principled and robust way to combine shape, pose and intensity features has been elusive for four main reasons: 1) heterogeneity of the data (data with linear and non-linear natural variation across features); 2) sub-optimal representation of three-dimensional Euclidean motion; 3) artificial discretization of the models; and 4) lack of an efficient transfer learning process to project observations into the latent space. This work proposes a novel statistical modelling framework for multiple bone structures. The framework provides a latent space embedding shape, pose and intensity in a continuous domain allowing for new approaches to skeletal joint analysis from medical images. First, a robust registration method for multi-volumetric shapes is described. Both sampling and parametric based registration algorithms are proposed, which allow the establishment of dense correspondence across volumetric shapes (such as tetrahedral meshes) while preserving the spatial relationship between them. Next, the framework for developing statistical shape-kinematics models from in-correspondence multi-volumetric shapes embedding image intensity distribution, is presented. The framework incorporates principal geodesic analysis and a non-linear metric for modelling the spatial orientation of the structures. More importantly, as all the features are in a joint statistical space and in a continuous domain; this permits on-demand marginalisation to a region or feature of interest without training separate models. Thereafter, an automated prediction of the structures in images is facilitated by a model-fitting method leveraging the models as priors in a Markov chain Monte Carlo approach. The framework is validated using controlled experimental data and the results demonstrate superior performance in comparison with state-of-the-art methods. Finally, the application of the framework for analysing computed tomography images is presented. The analyses include estimation of shape, kinematic and intensity profiles of bone structures in the shoulder and hip joints. For both these datasets, the framework is demonstrated for segmentation, registration and reconstruction, including the recovery of patient-specific intensity profile. The presented framework realises a new paradigm in modelling multi-object shape structures, allowing for probabilistic modelling of not only shape, but also relative pose and intensity as well as the correlations that exist between them. Future work will aim to optimise the framework for clinical use in medical image analysis.
  • Thumbnail Image
    Item
    Open Access
    Towards a framework for multi class statistical modelling of shape, intensity, and kinematics in medical images
    (2021) Fouefack, Jean-Rassaire; Burdin, Valérie; Mutsvangwa, Tinashe; Borotikar, Bhushan; Douglas, Tania
    Statistical modelling has become a ubiquitous tool for analysing of morphological variation of bone structures in medical images. For radiological images, the shape, relative pose between the bone structures and the intensity distribution are key features often modelled separately. A wide range of research has reported methods that incorporate these features as priors for machine learning purposes. Statistical shape, appearance (intensity profile in images) and pose models are popular priors to explain variability across a sample population of rigid structures. However, a principled and robust way to combine shape, pose and intensity features has been elusive for four main reasons: 1) heterogeneity of the data (data with linear and non-linear natural variation across features); 2) sub-optimal representation of three-dimensional Euclidean motion; 3) artificial discretization of the models; and 4) lack of an efficient transfer learning process to project observations into the latent space. This work proposes a novel statistical modelling framework for multiple bone structures. The framework provides a latent space embedding shape, pose and intensity in a continuous domain allowing for new approaches to skeletal joint analysis from medical images. First, a robust registration method for multi-volumetric shapes is described. Both sampling and parametric based registration algorithms are proposed, which allow the establishment of dense correspondence across volumetric shapes (such as tetrahedral meshes) while preserving the spatial relationship between them. Next, the framework for developing statistical shape-kinematics models from in-correspondence multi-volumetric shapes embedding image intensity distribution, is presented. The framework incorporates principal geodesic analysis and a non-linear metric for modelling the spatial orientation of the structures. More importantly, as all the features are in a joint statistical space and in a continuous domain; this permits on-demand marginalisation to a region or feature of interest without training separate models. Thereafter, an automated prediction of the structures in images is facilitated by a model-fitting method leveraging the models as priors in a Markov chain Monte Carlo approach. The framework is validated using controlled experimental data and the results demonstrate superior performance in comparison with state-of-the-art methods. Finally, the application of the framework for analysing computed tomography images is presented. The analyses include estimation of shape, kinematic and intensity profiles of bone structures in the shoulder and hip joints. For both these datasets, the framework is demonstrated for segmentation, registration and reconstruction, including the recovery of patient-specific intensity profile. The presented framework realises a new paradigm in modelling multi-object shape structures, allowing for probabilistic modelling of not only shape, but also relative pose and intensity as well as the correlations that exist between them. Future work will aim to optimise the framework for clinical use in medical image analysis.
  • Thumbnail Image
    Item
    Open Access
    Visualisation and manipulation of 3D patient-specific bone geometry using augmented reality
    (2020) Coertze, Johannes A; Mutsvangwa, Tinashe; Douglas, Tania; Bracio, Boris R
    Computer-mediated reality technologies have the potential to improve the imageguided surgery (IGS) workflow; specifically, pre-surgical planning, intra-operative guidance, post-surgical assessment, and rehabilitation. Augmented reality (AR), a form of computer-mediated reality, uses an electronic display or projection module to add a hologram in the user's field of view (FOV). For intra-operative guidance, AR could aid in reducing the cognitive overload experienced by clinicians due to integrating multi-modal imaging data from several sources while performing the intervention on the patient. Three AR HMD systems have been developed to explore the capabilities of the Microsoft HoloLens as an AR HMD to be used in developing an AR HMD medical system. The three AR HMD systems required different software and hardware system architectures, however, each of the AR HMD system's software applications has been developed in Unity combined with the Mixed Reality Toolkit (MRTK). Each of the AR HMD systems implemented different registration techniques to localize the virtual object in the real-world coordinate system. The registration techniques were user calibration alignment to identified anatomical landmarks, fiducial marker tracking, and markerless tracking. For user calibration with anatomical landmarks, the MRTK was manipulated to allow alignment of the virtual object. For fiducial registration, the Vuforia Software Development Kit (SDK) was added to assess the alignment and spatial anchoring of the virtual object as specified. Finally, the Leap Motion Controller (LMC) and Leap's Orion SDK was used for exploring markerless tracking. The AR HMD systems developed enabled performance assessments, and alignment errors were identified during trials of the three systems. Most notably the location drift of the 3D virtual object in the spatial space due to the clinician moving around the registered location. This project entailed preliminary development towards the AR HMD medical system to create an in-vivo view of 3D patient-specific bone geometries as a hologram in the clinician's FOV.