Use this URL to cite or link to this record in EThOS:
Title: Machine learning for efficient recognition of anatomical structures and abnormalities in biomedical images
Author: Xie, Zhongliu
ISNI:       0000 0004 6061 6712
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2017
Availability of Full Text:
Access from EThOS:
Access from Institution:
Three studies have been carried out to investigate new approaches to efficient image segmentation and anomaly detection. The first study investigates the use of deep learning in patch based segmentation. Current approaches to patch based segmentation use low level features such as the sum of squared differences between patches. We argue that better segmentation can be achieved by harnessing the power of deep neural networks. Currently these networks make extensive use of convolutional layers. However, we argue that in the context of patch based segmentation, convolutional layers have little advantage over the canonical artificial neural network architecture. This is because a patch is small, and does not need decomposition and thus will not benefit from convolution. Instead, we make use of the canonical architecture in which neurons only compute dot products, but also incorporate modern techniques of deep learning. The resulting classifier is much faster and less memory-hungry than convolution based networks. In a test application to the segmentation of hippocampus in human brain MR images, we significantly outperformed prior art with a median Dice score up to 90.98% at a near real-time speed (< 1s). The second study is an investigation into mouse phenotyping, and develops a high-throughput framework to detect morphological abnormality in mouse embryo micro-CT images. Existing work in this line is centred on, either the detection of phenotype-specific features or comparative analytics. The former approach lacks generality and the latter can often fail, for example, when the abnormality is not associated with severe volume variation. Both these approaches often require image segmentation as a pre-requisite, which is very challenging when applied to embryo phenotyping. A new approach to this problem in which non-rigid registration is combined with robust principal component analysis (RPCA), is proposed. The new framework is able to efficiently perform abnormality detection in a batch of images. It is sensitive to both volumetric and non-volumetric variations, and does not require image segmentation. In a validation study, it successfully distinguished the abnormal VSD and polydactyly phenotypes from the normal, respectively, at 85.19% and 88.89% specificities, with 100% sensitivity in both cases. The third study investigates the RPCA technique in more depth. RPCA is an extension of PCA that tolerates certain levels of data distortion during feature extraction, and is able to decompose images into regular and singular components. It has previously been applied to many computer vision problems (e.g. video surveillance), attaining excellent performance. However these applications commonly rest on a critical condition: in the majority of images being processed, there is a background with very little variation. By contrast in biomedical imaging there is significant natural variation across different images, resulting from inter-subject variability and physiological movements. Non-rigid registration can go some way towards reducing this variance, but cannot eliminate it entirely. To address this problem we propose a modified framework (RPCA-P) that is able to incorporate natural variation priors and adjust outlier tolerance locally, so that voxels associated with structures of higher variability are compensated with a higher tolerance in regularity estimation. An experimental study was applied to the same mouse embryo micro-CT data, and notably improved the detection specificity to 94.12% for the VSD and 90.97% for the polydactyly, while maintaining the sensitivity at 100%.
Supervisor: Gillies, Duncan ; Rueckert, Daniel Sponsor: China Scholarship Council ; Imperial College London
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral