Grant: Discovery Grants Program - Individual

As interventional imaging and advancements in Computer Science are incorporated into clinical workflows, computer-assisted interventions become increasingly reliable. Combining magnetic resonance (MR) imaging with a linear accelerator (Linac), the emerging MR-Linac technology enables to direct the radiating beam precisely at a moving tumor in external beam radiotherapy. Although effective to a degree, it is still impossible to track in `real-time' and in 3 dimensions (3D) the position of the treatment volume and surrounding organs at risk while concurrently delivering the radiation. In this research program, I propose to develop innovative methods for target tracking on MR images. This translational research will rely on accumulating knowledge in Computer Science to create a novel paradigm that addresses the gap created by the emerging MR-Linac technology. Acquisitions of 2D MR images are useful in capturing target(s) motion in nearly `real-time', whereas lengthy 3D image acquisitions are used to determine the reference position of the target for planning. Through paradigm shifts in image processing, artificial intelligence and predictive methods, the advanced methods aim to provide the `real-time' 3D displacement of the treatment target relative to its reference position. The first method to be developed, employs machine learning methods to track targets. The goal is to learn to extract the morphological signature of a target in a fixed image. That morphological signature is then used for detecting the corresponding target in the moving image. The second method to be developed aims to estimate the `real-time' position and shape of concurrent targets, which can be partially or fully unperceivable on images. To achieve this objective, surrogate measurements are used by a neural network to infer the position and shape of the targets. The third method to be developed relies on image registration. The mapping of common features between multi-modal 3D and 2D images is challenging. In turn, five sub-strategies will be researched. The first sub-strategy investigates the prediction of a 2D image, orthogonal to the acquired `real-time' one, using deformation vector fields for dual 2D-to-3D registration. The second sub-strategy consists of in-plane 2D-to-2D rigid registrations between the `real-time' image and several slices extracted from the reference volume. The third sub-strategy studies predictive methods for updating the target's out-of-plane displacement. The fourth sub-strategy mitigates appearance discrepancies between the 3D volume and `real-time' images by introducing an intermediate reference image. Finally, the fifth sub-strategy explores the 3D volume translation to resemble the `real-time' images, for subsequent image registrations. This research proposal will advance the image processing field. The primary application of this research is for computer-assisted interventions and applies to other fields, such as the automobile industry.