Deformation Estimation and Assessment of Its Accuracy in Ultrasound Images

This thesis aims to address two problems; one in ultrasound elastography and one in image registration. The first problem entails estimation of tissue displacement in Ultrasound Elastography (UE). UE is an emerging technique used to estimate mechanical properties of tissue. It involves calculating the displacement field between two ultrasound Radio Frequency (RF) frames taken before and after a tissue deformation. A common way to calculate the displacement is to use correlation based approaches. However, these approaches fail in the presence of signal decorrelation. To address this issue, Dynamic Programming was used to find the optimum displacement using all the information on the RF-line. Although taking this approach improved the results, some failures persisted. In this thesis, we have formulated the DP method on a tree. Doing so allows for more information to be used for estimating the displacement and therefore reducing the error. We evaluated our method on simulation, phantom and real patient data. Our results shows that the proposed method outperforms the previous method in terms of accuracy with small added computational cost. In this work, we also address a problem in image registration. Although there is a vast literature in image registration, quality evaluation of registration is a field that has not received as much attention. This evaluation becomes even more crucial in medical imaging due to the sensitive nature of the field. We have addressed the said problem in the context of ultrasound guided radiotherapy. Image guidance has become an important part of radiotherapy wherein image registration is a critical step. Therefore, an evaluation of this registration can play an important role in the outcome of the therapy. In this work, we propose using both bootstrapping and supervised learning methods to evaluate the registration. We test our methods on 2D and 3D data acquired from phantom and patients. According to our results, both methods perform well while having advantages and disadvantages over one another. Supervised learning methods offer more accuracy and less computation time. On the other hand, for bootstrapping, no training data is required and also offers more sensitivity.