Detection of anatomical structures in medical datasets
MetadataShow full item record
Detection and localisation of anatomical structures is extremely helpful for many image analysis algorithms. This thesis is concerned with the automatic identiﬁcation of landmark points, anatomical regions and vessel centre lines in three-dimensional medical datasets. We examine how machine learning and atlas-based ideas may be combined to produce efﬁcient, context-aware algorithms. For the problem of anatomical landmark detection, we develop an analog to the idea of autocontext, termed atlas location autocontext, whereby spatial context is iteratively learnt by the machine learning algorithm as part of a feedback loop. We then extend our anatomical landmark detection algorithm from Computed Tomography to Magnetic Resonance images, using image features based on histograms of oriented gradients. A cross-modality landmark detector is demonstrated using unsigned gradient orientations. The problem of brain parcellation is approached by independently training a random forest and a multi-atlas segmentation algorithm, then combining them by a simple Bayesian product operation. It is shown that, given classiﬁers providing complementary information, the hybrid classiﬁer provides a superior result. The Bayesian product method of combination outperforms simple averaging where the classiﬁers are sufﬁciently independent. Finally, we present a system for identifying and tracking major arteries in Magnetic Resonance Angiography datasets, using automatically detected vascular landmarks to seed the tracking. Knowledge of individual vessel characteristics is employed to guide the tracking algorithm by two means. Firstly, the data is pre-processed using a top-hat transform of size corresponding to the vessel diameter. Secondly, a vascular atlas is generated to inform the cost function employed in the minimum path algorithm. Fully automatic tracking of the major arteries of the body is satisfactorily demonstrated.