Reconstruction, localisation, and segmentation in medical images

Abstract

Medical imaging plays a crucial role for identifying a medical condition, therapeutic response, disease level or the chance of developing a future complication. This thesis proposes robust and fully-automatic medical imaging methods that can assist doctors in their clinical decision making. In particular, it contributes to the main components of the medical imaging pipeline: image acquisition, reconstruction, and analysis. A novel approach is proposed for motion correction and reconstruction in foetal MRI. Whereas scanned images suffer from inter-slice motion artefacts because of the unconstrained foetal movements. The proposed method explores a new reconstruction paradigm that splits the 3D input image into overlapping patches. This enables conventional reconstruction methods to compensate for non-rigid deformations and reconstruct the whole uterus. Another new method is proposed for automatic foetal brain localisation, which is essential for assessing the brain development and maturation in foetal MRI. The proposed method uses superpixel graphs to train a random forest classifier, which allows to efficiently extract 3D global features instead of 2D local features. Results show high brain detection accuracy, without the need of any landmarks or prior information such as the gestational age of the foetus. Automatic segmentation of the placenta is more challenging than brain segmentation because of its variability in position, shape and appearance. The proposed method adopts a multi-pathway convolutional neural network to efficiently captures both local and global contextual information. The predicted output is then refined using a condition random field to regularise outliers. This thesis also proposes and evaluates reinforcement learning based methods for the detection of anatomical landmarks, which is an important step for a wide range of applications in medical image analysis. Manual annotation of such landmarks is a tedious task and prone to observer errors. Fixed-scale and multi-scale search strategies are evaluated on three different datasets: foetal ultrasound, brain and cardiac MRI. The evaluation results outperform current supervised and other reinforcement learning state-of-the-art methods. Finally, a novel approach is proposed for automatic view planning using intelligent artificial agents to mimic the navigation process towards the target plane. Similar to the landmark agents, view planning agents learn to find the target plane iteratively using reinforcement learning. The experiments show that view planning agents can detect different standard planes from brain and cardiac MRI with state-of-the-art accuracy in a very short time.