Master Thesis AI Based Prediction for Anatomical Motion Modelling on Unity

Elekta

Stockholm · Posted 3h ago

Apply

At a glance

Location: Stockholm, Sweden
Work type: Onsite

Master thesis project at Elekta in Stockholm focusing on AI-based prediction of anatomical motion for Unity-based simulation, investigating how AI methods can improve motion prediction horizons within the APM framework for MR-Linac radiotherapy.

What they are looking for

PythonAiDeep learningMedical image processingProgramming

Location: Stockholm

Are you a current Elekta employee?

Please click here to apply through our internal career site Find Jobs - Elekta.

We don’t just build technology. We build hope for everyone dealing with Cancer.

Elekta is shaping the standard of care, together with healthcare providers, by developing precision radiation medicine that brings advanced radiotherapy into routine clinical practice. Our solutions support consistent, efficient and personalized care, addressing evolving patient needs. Through adaptive radiotherapy, precision treatment delivery and integrated workflows, Elekta supports care teams across a range of clinical settings. Each year more than 2 million patients are treated using Elekta solutions across over 130 countries. Elekta is headquartered in Stockholm, Sweden, with over 4,000 employees and offices in more than 40 countries. The company is listed on Nasdaq Stockholm. For more information, visit elekta.com.

Introduction

Radiotherapy targets and surrounding anatomy can move during treatment. The Elekta Unity combines an MR scanner with a linear accelerator, enabling anatomy to be visualised during radiation delivery and allowing treatment adaptation based on real-time information. However, there are inherent latencies associated with image acquisition, image processing, data transfer, and motion estimation. To compensate for these latencies, future anatomical motion must be predicted. This project will investigate whether an AI-based approach can improve prediction accuracy at extended prediction horizons.

Background

Radiotherapy treatments are designed with millimeter precision to ensure that high doses of radiation can be delivered to the target while sparing surrounding healthy tissues and critical anatomical structures. However, both the target and surrounding anatomy may move during treatment due to physiological processes such as respiration, peristalsis, and organ drift. The Elekta Unity can visualise and respond to this motion in real time, improving dosimetric accuracy and therefore improving patient outcomes.

The current Automatic Position Management (APM) framework acquires an orthogonal pair of 2D cine images at a rate of 5 Hz. Each image is compared with a reference image using rigid registration to estimate target motion. A prediction engine then extrapolates the observed motion to estimate the target position once all processing steps have been completed.

In the future, longer prediction horizons may be required to support more advanced image acquisition strategies, target-tracking approaches, or dose-based interventions. To enable these developments, a prediction algorithm must maintain accuracy while predicting further into the future. AI methods such as recurrent neural networks have shown promise for time-series prediction tasks both within and beyond radiation oncology and will therefore form the primary focus of this investigation.

Problem Statement

The project will investigate the suitability of AI approaches for motion prediction on the Unity platform through the following objectives:

Curate a dataset of target centroid motion from a range of anatomies influenced by respiratory motion using a combination of publicly available datasets and historical MR-Linac treatment data.

Establish a benchmark by evaluating the curated dataset using the currently implemented prediction algorithm.

Choose a suitable AI model to predict centroid motion at clinically relevant prediction horizons. Adapt the model and train on the data collected.

Compare the performance of the proposed approach against the existing prediction algorithm using appropriate quantitative metrics.

Output

An assessment of the potential benefits of incorporating AI-based motion prediction within the APM framework

A curated motion dataset that can support future research and development activities.

Prototype software, reproducible processing pipelines, and documentation.

A written report and final presentation

Who are we looking for?

We are looking for a master’s student in medical physics, computer science, applied mathematics, engineering physics, or a related scientific discipline.