Search

Background: Individuals with Borderline Personality Disorder [BPD] symptomatology have high rates of self-harm (50-80%). Limited information exists on the most appropriate interventions to prevent recurrent self-harm in this population. Recent reviews on BPD interventions have suggested more research needs to be conducted looking at how individuals experience interventions with the aim of identifying the effective components of interventions. Objectives: The aim of the present study was to examine how individuals with BPD symptomatology experience interventions to prevent self-harm using a qualitative methodology. Methodology: Twelve individuals with BPD symptomatology and past or current self-harm were recruited through therapeutic services, and took part in a semi-structured interview. The interviews were analysed using a grounded theory approach. Findings: The grounded theory identified a core category, an alternative path to self-harm, and two sub-categories, established beliefs and causal factors, and the time course of self-harm. The results were presented using a process model which was indicative of the participants' experiences of interventions. Conclusions: The findings suggest individuals with BPD symptomatology perceive interventions as helping to reduce self-harm when interventions are long-term, consistent, and instant, and the intervention's outcome matches the purpose for the self-harm. The use of interventions appears to be context dependent, specifically being affected by the individual's level of emotional tension, and their cognitive processing during the decision to seek help. For long-term self-harm prevention, multiple interventions are required, and individuals need to be actively maintaining and evaluating these alternative strategies. It is suggested adoption of such a holistic approach could be one avenue for developing collaborative and effective self-harm interventions in clinical practice.

Three-dimensional (3D) ultrasound imaging has contributed to our understanding of fetal developmental processes in the womb by providing rich contextual information of the inherently 3D anatomies. However, its use is limited in clinical settings, due to the high purchasing costs and limited diagnostic practicality. Freehand two-dimensional (2D) ultrasound imaging, in contrast, is routinely used in standard obstetric exams. The low cost and portability of 2D ultrasound render it uniquely suitable for use in low- and middle-income settings. However, high level of expertise is always involved and it inherently lacks a 3D representation of the anatomies, which limit its potential for more accessible and advanced assessment. Capitalizing on the flexibility offered by freehand 2D ultrasound acquisition, this thesis presents a deep learning-based framework for optimizing the utilization and diagnostic power of 2D freehand ultrasound in fetal brain imaging. First, a localization model is presented to predict the location of 2D ultrasound fetal brain scans in the 3D brain atlas. It is trained by sampling 2D slices from aligned 3D fetal brain volumes, such that heavy annotations for each 2D scan are not required. This can be used for scanning guidance and standard plane localization. An unsupervised methodology is further proposed to adapt a trained localization model to freehand 2D ultrasound images acquired from arbitrary domains, for example sonographers, manufacturers and acquisition protocols. This enables the model to be used at the bedside in practice, where it can be fine-tuned with just the images acquired in any arbitrary domains before inference. Building upon the ability to localize 2D scans in the 3D brain atlas, a framework is further presented to reconstruct 3D volumes from non-sensor-tracked 2D ultrasound images using implicit representation. With this slice-to-volume reconstruction framework, additional 3D information can be extracted from the 2D freehand scans. Finally, a semi-automatic model, trained only on raw 3D volumes without any manual annotation, is presented to segment any arbitrary structures of interest in 3D medical volumes, while only requiring manual annotation of a single slice during inference. The model is tested on wide variety of medical imaging datasets and anatomical structures, verifying its generalizability. In the design of the framework presented in this thesis, three fundamental principles, namely minimal human annotation, generalizability and sensorless operation, are followed to optimize its seamless integration into the clinical workflow. This may modernize freehand routine scanning and enhance its accessibility, while maximizing the clinical information gained from routine scans acquired as part of the continuum of pregnancy care.

Gestational age estimation is a key marker in obstetric care for determination of fetal growth and health. Current clinical estimation of gestational age is based on regression using clinical fetal growth charts and biometry measurements by an experienced sonographer. However, estimation of gestational age using biometry is difficult especially in late pregnancy, and there may be additional salient anatomical information in images to assist gestational age estimation in standard planes which is discarded by clinical biometry measurement. This thesis therefore attempts to use deep neural network to directly regress gestational age from fetal standard planes without pixel size information, thus relying on anatomical appearance alone without fetal biometry. In this thesis, we first investigate the data augmentation methods commonly used for medical image analysis and implement an data augmentation selection strategy. We find that this optimized augmentation strategy leads to increased performance on the proxy task of standard plane classification compared to conventional hand-crafted data augmentation methods. We then use this data augmentation strategy selection framework to train a deep convolutional neural network for gestational age estimation from the head, abdomen and femur standard planes, and investigate the performance of different loss functions for deep regression. We demonstrate gestational age estimation across the second and third trimester on the head standard plane with a mean absolute error of 0.6 weeks with the optimized data augmentation policy, and use the trained neural network to visualize anatomically salient areas of the input image. We then extend the deep learning framework by introducing priors to trainable weights, making a Bayesian neural network. The Bayesian neural network separately estimates aleatoric and epistemic uncertainty during gestational age inference. This allows for calibrated uncertainties during gestational age estimation. This is the first work to estimate fetal gestational age from ultrasound standard plane images without pixel size or biometry information, and may be potentially of use in low- and medium- income countries where accurate gestational age dating in late pregnancy is important as women present to obstetric care in all stages of pregnancy.

One of the current challenges in applying machine learning to medical images is the difficulty in obtaining labelled training data. While medical images themselves are often available, generating high-quality training labels for them is time-consuming and often requires a trained clinician. This problem is particularly acute in 3D segmentation, which generally requires detailed voxel-by-voxel segmentation maps. This thesis proposes a series of image analysis methods to leverage additional available information to improve image segmentation quality when there are limited manual labels available. Using a dataset of 3D fetal ultrasound scans with only a small number of initial segmentation labels, we demonstrate a benefit from different information sources. The approaches used in this thesis are: • Generating segmentation labels automatically from a publicly available spatio-temporal atlas, making use of prior anatomical information. We use anatomical keypoints to guide registration of atlases across different imaging modalities. We then propagate these segmentations to individual ultrasound volumes, automatically generating a set of multi-label segmentations with an average Dice coefficient of 0.818 across the chosen anatomical structures in the range of 20-25 gestational weeks. This set of labels is used to train a segmentation CNN. We find that this produces high-quality segmentations, and that training a single network with the full multi-label training set improves segmentation quality (average Dice coefficient of 0.726 across structures for a multi-task network, versus 0.659 for individual single-task networks). • Leveraging additional, unlabelled data to improve performance for a cerebellar segmentation task. Unlabelled data does not have a ground-truth label by which segmentation quality may be compared, but measures of output uncertainty can be derived using test-time augmentation and dropout measures. We use these estimates of uncertainty to inform an iterative omni-supervised framework, using the highest-quality segmentations from the unlabelled dataset as additional training data for a segmentation CNN. We show that the use of these labels as part of a training dataset improves segmentation performance, and that this is improved further when measures of uncertainty are used in data selection (an overall improvement in Dice coefficient for a cerebellar segmentation task from 0.673 to 0.727). • Passing additional, scalar information to a neural network's input. We propose an original method by which scalar data such as gestational age, readily available as part of clinical data acquisition, may be appended to the input of a fully-convolutional neural network. We show that including this information as part of the training dataset can lead to significant improvements in segmentation quality on the same cerebellar segmentation task, in the reported experiments improving test Dice coefficient from 0.673 to 0.723. We find that this improvement is not sensitive to imperfect information passed at test time. The computational methods developed in this thesis may reduce the reliance on large manually-labelled datasets for medical image analysis and save time in generating expert annotations.

Obstetric ultrasound has proven an integral part of prenatal care for many applications including detecting pregnancy risk factors and fetal abnormalities. Although ultrasound imaging is widely used in developed countries, its availability in limited-resource settings has been dependent on device portability, associated costs, and trained healthcare workers. In recent years, as ultrasound equipment has become cheaper and portable its availability in limited-resource settings has expanded with promising trends. Yet a major constraint of ultrasound imaging is the steep learning curve to reach scanning proficiency, rendering the process challenging for inexperienced users. This thesis presents methods towards addressing the fundamental task of empowering novice sonographers to perform targeted diagnostic screening to the same level as experienced ultrasonographers. This is of particular relevance to low-resource settings where pregnant women may not be able to attend specialist obstetric care centres during their pregnancy. We address this task by introducing simplified scanning protocols and using computer vision and machine learning techniques. While aimed at low-income country healthcare needs, the research is also equally applicable to needs of developed world. Firstly, an original framework for automatic detection of frames of interest in ultrasound videos is presented. Furthermore, the merits of pre-processing images on classification accuracy, on datasets of variable size, is investigated. To this end, a classification framework for 2D ultrasound frames based on the improved Fisher vector encoding of SIFT and LP-SIFT features has been introduced. Additionally a Structured Random Forests edge detection framework is introduced for an accurate structure representation in 2D ultrasound data. Secondly, a novel framework based on video dynamics and motion patterns is presented for automatic detection of sequences of interest in real world ultrasound videos. To illustrate clinical applicability, this approach is applied to automatically identify the fetal presentation in the womb, during the third trimester of pregnancy from a predefined freehand ultrasound video sweep. A kernel dynamical textures model is utilised to model the dynamics of video sub-sequences that contain a fetal skull and those sequences that do not. A similarity kernel is then constructed between the model parameters to train a classifier. A classification accuracy of 93% was achieved using this framework. Thirdly, for the first time, the real world application of automatic analysis of predefined ultrasound videos to detect the fetal presentation and heartbeat is considered. For this application, the frames of interest are initially identified using a frame-by-frame classification framework. Moreover, a linear-chain structured graphical model is used to take into account the temporal information in the 2D predefined sweeps and to regularise the classification results. Consecutive frames of the fetal heart are further investigated for the presence of the heartbeat activity. Classification accuracy achieved using our method was about 88%. Finally, potential future directions are investigated and initial findings have been reported. These include results on a small pilot study specifically designed to perform a statistical comparison between images obtained from a standard mid-range ultrasound probe and a low-end portable ultrasound probe. Moreover, the efficacy of using a deep convolutional neural network with transfer learning is investigated and initial findings are reported.

Joint kinematics has been suggested to play an important role in the initiation and progression of a number of musculoskeletal pathologies. Thus, investigation of joint kinematics can help clinicians and researchers to better understand musculoskeletal conditions such as osteoarthritis and trochanteric bursitis. Existing motion capture technologies used in clinical settings suffer from various limitations, including soft tissue artefact. These limitations result in a lengthy examination process, an incomplete movement description, or an inaccurate representation of the real bony landmarks. The recent introduction of a motion analysis with ultrasound (MAUS) system aimed to provide a less constrained approach to detect actual bony structures and describe joint kinematics in three dimensional (3D) space. However, the accuracy of the original MAUS system was highly dependent on the operator's experience. In this doctoral thesis, a computer-aided tracking and motion analysis with ultrasound (CAT & MAUS) system is developed to track underlying bony landmarks and describe hip joint kinematics during gait. The key contribution of this thesis is to combine state-of-the-art computer vision approaches with the original MAUS system to improve the speed, accuracy and repeatability of joint kinematics examination for clinical measurement and diagnosis. Firstly, a comprehensive review of gait analysis and relevant clinical diagnostic modalities is presented. Then, an augmented MAUS system architecture is presented, which is more flexible than the previous MAUS system of the data acquisition. It combines an optoelectronic motion analysis (MA) system with a 2D ultrasound (US) device to build up a 3D representation of the bony structure of interest. A novel calibration box with multiple functions for the augmented MAUS system is designed to spatially and temporally match US images to the motion analysis data. The average Euclidean distance error of the spatial calibration is found to be 0.34 mm and the accuracy of the temporal calibration is found to be within half of the frame acquisition interval. The augmented MAUS system with its more accurate calibration procedures can accurately present the bone of interest in 3D space. Secondly, a computer-aided post-processing pipeline is presented to automatically track the bone of interest in 3D space. A globally optimal registration method is employed to align the 3D surfaces of the target bony structure reconstructed from the manual segmentation in US images at different positions to locate the target bony landmark from one position to another. The globally optimal registration overcomes the issue of getting trapped into local minima for the conventional iterative closest point registration. The accuracy of the globally optimal registration is validated with a proximal femur phantom. The average rotation error is found to be 0.38° and the average translation error is found to be 0.33 mm, both of which are within the clinical tolerance of computer-aided orthopaedics surgery ±3° and ± 1 mm). The 3D globally optimal registration guarantees an accurate track of the same target bony structure at different positions instead of estimating the similar bony structure at each position by eyeball with previous MAUS system. However, in practice, it takes around 20 minutes to manually segment the bone structure for a single surface reconstruction, which is extremely laborious and time consuming. In order to automate the CAT & MAUS system, a novel automatic segmentation of the bone structure in a 2D US image is then developed as a precursor for an accurate globally optimal registration. The automatic bone segmentation introduces the local phase features and acoustic characteristics to enhance the bone probability for detection. The result of the automatic segmentation is validated with a pilot phantom study before applying to the in-vivo experiment and achieves an accuracy of 0.13 mm. After fully developing the CAT & MAUS system, hip joint kinematics of healthy subjects are quantified in six degrees of freedom using the CAT & MAUS system and compared to the results from the optoelectronic motion analysis system. It is shown that CAT & MAUS results describe a greater rotation range than MA results by up to 4.03° in the sagittal plane because the optoelectronic motion analysis system alone suffers from severe soft tissue artefact as the skin markers shift away from the underlying bony structures during movement. The deformation of body segments captured by MA during movements caused by soft tissue artefact is explained using Procrustes analysis to indicate the accuracy and repeatability of the CAT & MAUS system. Finally, possible future directions are proposed based on preliminary investigations using the developed CAT & MAUS system with particular reference to both the technical and clinical perspectives.

The aim of this thesis was to develop novel embolic particles for cancer therapy using a combination of nano and microparticles. The particles have been designed and synthesised to have a polystyrene core, be radiopaque, act as a radiosensitiser and have a high capacity for loading of chemotherapy drugs. The polystyrene core has been designed so it effectively blocks the tumour vasculature thereby limiting oxygen and nutrient delivery to the tumour thus causing tumour necrosis. The core of the polystyrene particle incorporates tantalum oxide nanoparticles to provide X-ray contrast to the embolic microparticles. The surface of the polystyrene particle has been coated with rare earth doped titanium dioxide (TiO2) nanoparticles which produce reactive oxygen species upon X-ray activation. This allows the embolic particle to act as a radiosensitising agent and has been shown to reduce cell proliferation in the presence of X-rays. The surface of the polystyrene particles has also been coated with mesoporous SiO2 nanoparticles which allow for the high loading capacity of chemotherapeutic drugs. This permits chemotherapy to be delivered directly to the tumour location, thereby reducing the toxic side effects of systemic treatment. As a proof of concept, the chemoembolization particles have been loaded with a novel fungal derived chemotherapeutic Ophiobolin A and the controlled release has been demonstrated. The OphA chemoembolization particles reduced cell viability by approximately 70% compared to the blank chemoembolization particles. The mechanism of cell death of OphA on eight cancer cell lines and one control cell line has also been studied and its effect on cellular organelles elucidated. OphA shows great promise as a novel chemotherapeutic which could be taken forward to animal trials.

In the research of cancer biomarker and cell signaling pathways, multiple immunofluorescence labelled tissue microarrays (TMAs) are often examined by pathologists. By looking at TMAs under microscopy, the pathologists can score the images visually. In order to improve the accuracy and reliability of the image scoring, an automatic quantification of TMA image is in high demand. This thesis presents the development and the validation of several key components of image analysis in order to automatically quantify TMA images. The first component is preprocessing of the image, the uneven illumination in fluorescent confocal microscopic images are corrected using a retrospective method and it is validated on this dataset. The second component is the design and validation of an original cell segmentation algorithm. This algorithm automatically segments cell nuclei and cell cytoplasm from the images. The validation of this algorithm shows more than 90% of accuracy subject to tissue images with good qualities. The quality of tissue images refers to the problems such as, clumpy and overlaping nuclei, DNA degradation, stromal cells and etc. As the tissue quality varies, the texture of tissue changes. The texture of tissue with different quality was analyzed and a classification model is built to classify the textures. The third component is an automatic tissue quality assessment algorithm for TMA images. In addition, an original framework is designed to assess the quality of TMA by automatically recognize different tissue types and produce the proportion of each type. This proportion of tissue types provides a quality metric for an image and it is shown in the study that this metric is useful for classify the tissue image quality. In conclusion, this thesis presents an original the pipeline including the three components, to analyze tissue microarray images. The validation of each method has shown good performance.

Global assessment of myocardial function is widely performed by estimating ejection fraction (EF), but many common cardiac diseases initially affect the myocardium on a regional, rather than global, basis. Computed tomography (CT), most commonly applied to assess the coronary arteries, is a prime candidate for such regional analysis. This doctoral thesis makes steps towards regional CT functional analysis with two clinical and two technical contributions. The first clinical contribution focuses on evaluating the feasibility and utility of functional analysis with currently available CT technology. Our study found that CT strain analysis could identify regional wall motion abnormalities in cardiomyopathy that are not otherwise detected using conventional metrics of myocardial function such as EF. In order for cine CT of the heart to become routine clinical practice, improvements need to be made to the image acquisition protocol. The second clinical contribution focus on making these improvements with results pointing to the possibility of one millisievert range cine CT images with high (>50 milliseconds) temporal resolutions. Moving to technical considerations, a key concern has been how to better characterise the myocardium in CT. To address this, the first technical contribution examines the use of feature-based attribute vectors, which were found to improve image registration towards deriving more reliable motion estimations. The second technical contribution focus on developing a pipeline tailored towards CT strain analysis. Noting that CT naturally provides information in 3D, a 3D hyperelastic biomechanical model fitting method was evaluated. Analysis of an infarction model demonstrated that regional myocardial strain can be estimated in the 3D space and areas of infarction can be detected. By considering both technical and clinical perspectives, these advances will contribute to the the field of regional cardiac functional analysis towards improving patient care.