Master Thesis Scatter Estimation in Computed Tomography

Master Thesis Scatter Estimation in Computed Tomography

Introduction

Cone beam computed tomography (CBCT) is a method of forming a 3D image from a set of X-ray projections. The projections are formed by exposing a subject to a cone-shaped X-ray beam and capturing the photons in a flat-panel detector, similar to the detector of a camera. The subject both attenuates and scatters the X-ray beam, affecting what photons are detected by the detector. Projections are formed from multiple directions. This results in an inverse problem that can be solved to reconstruct a 3D image of the subject.

There are efficient algorithms to solve this inverse problem when there is no scattering (FDK or iterative solvers). A common approach to CBCT reconstruction is to estimate the scattering and effectively subtract it from projection images so that these efficient algorithms can be employed for reconstruction.

Background

At Elekta, we develop products for cancer radiotherapy that use CBCT for imaging and positioning the patient. We are today using an AI model to estimate the scatter component in CBCT images. As an alternative to this, we have developed a physics-based algorithm (patent pending) for estimating scatter. We have made a prototype implementation of this algorithm in 2000 lines of Rust that we call Rainbow.

Problem Statement

There are several open questions related to our scatter estimates that could be the basis of a master's thesis. What we decide on depends on your interests and our needs.

• Engineering focus: How do we best evaluate Rainbow’s performance? What subjects do we scan, what metrics do we measure, how do we compare with state-of-the-art AI models? What applications is it suitable for?

• Physics focus: What improvements can be made to the physical model of Rainbow? We have made several approximations of the scattering and detector physics that can be improved for a more accurate model. This may involve both physics theory and experimental work in our labs.

• Implementation focus: To be clinically useful, Rainbow needs to run faster. Are there optimizations we can make to the current implementation? How do we efficiently implement this algorithm on a GPU?

• Math focus: What improvements can be made to the algorithm, such as Monte Carlo variance reduction or methods for accelerating convergence of a numerical solver.

Depending on our findings until the start of the master thesis work, we may have other topics related to CBCT reconstruction and scatter estimates.

Output

The student is expected to write a report that helps us understand the above-mentioned aspects of Rainbow. In the case of investigating various improvements, this should be accompanied by code contributions to Rainbow. In the case of benchmarking Rainbow, the report should be accompanied by scripts for running these benchmarks.

Who are we looking for?

Preferred Masters education: Physics or Mathematics

Required skills: Python programming, good understanding of statistics and numerical methods.

Useful skills: X-ray lab work, Rust programming, GPU computing, Monte Carlo simulations

Practicalities

Location: Forskaren, Hagaplan 4, Stockholm

Start date: Fall 2026 or early spring 2027