Session Item

Adaptive image-guided radiotherapy (IGRT) in the pelvis is often complicated by the large displacements introduced by bladder and rectal filling. While deformable image registration (DIR) can be employed, state-of-the-art DIR algorithms generally have difficulties in estimating such large-scale deformations. In this study we propose a DIR algorithm, adopting concepts from fluid mechanics, in order to address the large-scale deformations occurring during IGRT in the pelvis.

We propose addressing the registration problem by solving the following partial differential equation with respect to u:

∂SSD[u]/∂u + μ∆v + (μ + λ)∇(∇ · v) = 0,

where u is the deformation vector field (DVF), SSD is the sum of squared differences between the images, v = ∂u/∂t + v · ∇u is the velocity field, and μ, λ are two regularisation parameters. In brief, the ∂SSD/∂u term minimises the misalignment between the two images, whereas the remainder of the equation imposes the fluid behaviour for the DVF’s. An iterative solver has been implemented on the GPU using the CUDA programming library.

The accuracy of the proposed optical flow fluid regularisation (OFFR) model is evaluated on ten prostate cancer patients. Each patient had a series of five 3D-MR images acquired, with the first image in each series playing the role of anatomical reference. Clinical contours of the prostate, bladder and rectum were available for each image. The estimated DVF’s are used to propagate the contours, which are evaluated in terms of the Dice similarity coefficient and the Hausdorff distance.

Additionally, the performance of the proposed model was compared to several existing state-of-the-art DIR solutions: Thirion’s demons algorithm, Elastix and the Horn-Schunck method.

An overview of the scores on the bladder, prostate and rectum contours can be found in Fig. 1. The largest improvement of OFFR over the other algorithms is the alignment of the bladder. This is further exemplified in Fig. 2. The sagittal slice displays the significant initial volumetric dissimilarity in the bladder (Dice: 0.56, Hausdorff: 4.91 mm). OFFR resolves this problem by expanding the contours (Dice: 0.97, Hausdorff: 0.08 mm).

The proposed DIR algorithm was analysed for its performance to account for large anatomical deformations during IGRT in the pelvis. MR images of prostate cancer patients with delineations of the prostate, bladder and rectum were chosen. The results show an overall improvement in the propagation of the contours, particularly within the bladder, compared to several well-established DIR algorithms.