Medical Image Registration by Neural Networks A Regression-Based Registration Approac

This thesis focuses on the development and evaluation of a registration-by-regression approach for the 3D/2D registration of coronary Computed Tomography Angiography (CTA) and X-ray angiography. This regression-based method relates image features of 2D projection images to the transformation parameters of the 3D image by a nonlinear regression. It treats registration as a regression problem, as an alternative for the traditional iterative approach that often comes with high computational costs and limited capture range. First we presented a survey of the methods with a regression-based registration approach for medical applications, as well as a summary of their main characteristics (Chapter 2). Second, we studied the registration methodology, addressing the input features and the choice of regression model (Chapter 3 and Chapter 4). For that purpose, we evaluated different options using simulated X-ray images generated from coronary artery tree models derived from 3D CTA scans. We also compared the registration-by-regression results with a method based on iterative optimization. Different image features of 2D projections and seven regression techniques were considered. The regression approach for simulated X-rays was shown to be slightly less accurate, but much more robust than the method based on an iterative optimization approach. Neural Networks obtained accurate results and showed to be robust to large initial misalignment. Third, we evaluated the registration-by-regression method using clinical data, integrating the 3D preoperative CTA of the coronary arteries with intraoperative 2D X-ray angiography images (Chapter 5). For the evaluation of the image registration, a gold standard registration was established using an exhaustive search followed by a multi-observer visual scoring procedure. The influence of preprocessing options for the simulated images and the real X-rays was studied. Several image features were also compared. The coronary registration--by-regression results were not satisfactory, resembling manual initialization accuracy. Therefore, the proposed method for this concrete problem and in its current configuration is not sufficiently accurate to be used in the clinical practice. The framework developed enables us to better understand the dependency of the proposed method on the differences between simulated and real images. The main difficulty lies in the substantial differences in appearance between the images used for training (simulated X-rays from 3D coronary models) and the actual images obtained during the intervention (real X-ray angiography). We suggest alternative solutions and recommend to evaluate the registration-by-regression approach in other applications where training data is available that has similar appearance to the eventual test data.