Deep learning-based metal and scatter artifact reduction in conebeam computed tomography

Abstract

Cone-Beam Computed Tomography (CBCT) provides high-quality three-dimensional X-ray imaging and offers advantages such as reduced radiation dose, lower cost, and a smaller physical footprint compared to Multi-Detector CT (MDCT). Due to these features, CBCT is well-suited for a range of clinical applications, including dentistry, orthopedics, interventional radiology, and image-guided therapies, as well as for use in mobile clinics and remote deployments, thereby contributing to broader accessibility in healthcare. However, CBCT image quality is often compromised by inherent artifacts, including scatter and metal artifacts, which pose significant challenges to diagnostic applications.

The emergence of deep learning methods presents a promising avenue for addressing these imaging challenges, potentially offering substantial improvements. However, practical constraints, such as the need for large training datasets, and the seamless integration of deep learning models into existing artifact correction pipelines, must be addressed to ensure clinical feasibility.

The main purpose of this thesis is to develop clinically applicable deep learning techniques to mitigate metal and scatter artifacts in CBCT imaging. For instance, to overcome the challenge of data scarcity, simulated datasets are leveraged for network training. Additionally, lightweight Convolutional Neural Network (CNN) models are introduced to facilitate efficient integration into established artifact correction workflows. To ensure real-world applicability, the proposed methods are evaluated using real CBCT datasets. The research is structured around four key contributions. Publication I introduces a learning-based inpainting method of metal traces to reduce metal artifacts. Publication II employs simulated data to train a neural network for metal trace segmentation to improve the effectiveness of existing inpainting-based metal artifact reduction methods. Publication III presents a neural network for scatter estimation under clinically relevant variations in the Field of Measurement (FOM). Finally, in Publication IV, an ultrafast scatter estimation approach is proposed for deployment in mobile CBCT systems and on-device applications.

The findings demonstrate that the developed models substantially enhance the state-of-the-art in artifact correction, advancing the clinical viability of CBCT imaging through deep learning-driven lightweight solutions.