Improving precision of deformable image registration

Deformable image registration (DIR) has various applications in medical image analysis such as in adaptive radiotherapy (ART) and multi-atlas segmentation. ART uses DIR to warp the dose from the daily cone-beam CT (CBCT) image to the planning computed tomography (CT) image to accommodate anatomical changes that occur between dose planning and dose delivery. DIR aims to find a geometric transformation between corresponding image data and brings them into a common coordinate frame. DIR is an inherently ill-posed problem that lacks a unique mapping between the voxels of the two images being registered. As such, one must regularize the registration to achieve physically meaningful transforms. The regularization penalty is usually a function of derivatives of the displacement vector field and can be calculated either analytically or numerically. The numerical approach, however, is computationally expensive depending on the image size, and therefore a computationally efficient analytical framework has been developed. Cubic B-splines were used to model the displacement vector field. A generalized mathematical framework was developed that supports five distinct regularizers: diffusion, curvature, linear elastic, third-order, and total displacement. The approach was validated by comparing each regularizer with its numerical counterpart in terms of accuracy. Benchmarking results show that the analytic solutions run significantly faster - up to two orders of magnitude - than finite differencing-based numerical implementations. Additionally, the cost function used in the DIR optimization process is a global, low-level cost function. It fails to accurately align high-level image structures such as organs at risk (OARs). DIR accuracy can be improved locally by adding high-level structures such as contours and landmark points. A hybrid image similarity metric that incorporates a point-to-distance map (PD) metric was developed. Given a pair of segmented images, structures on the fixed image are represented as sets of points while structures on the moving image are described as distance maps. The total distance of all fixed points to their associated boundaries on the moving image constitutes the PD metric, which is combined with the more traditional intensity similarity metric between the fixed and moving images. The approach was validated using the pelvic reference dataset wherein the prostate, bladder, and rectum were manually segmented from the CT and CBCT images by a medical expert to obtain the segmented fixed and moving images. The accuracy of the deformable registration was quantified using the Dice Similarity Coefficient (DSC) and 95% Hausdorff Distance (95% HD), with and without the PD metric. Results demonstrate a much-improved overlap between the fixed and warped contours once the PD metric was applied. Moreover, the computational overhead associated with adding the PD metric was minimal. Finally, to circumvent the need for a medical expert to manually segment the structures, a 3-D U-Net architecture was used to generate structures to be used with the PD Metric calculations. An image-to-image translation network is trained using a CycleGAN to domain adapt between CT and CBCT images to assist the auto-segmentation process. Experiments are performed to compare the DIR accuracy with the 3-D U-Net generated structures and the manually segmented structures. Results demonstrate that the auto-segmented images improve the DIR accuracy equally when using manually segmented images. This helps in reducing manual overhead associated with segmenting the images and thus increases the end-to-end performance of the DIR algorithm making it suitable for clinical use.