Towards engineering ultrasound image analysis solutions for resource constrained environments

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.