CT metal artifact reduction with machine learning and photon-counting techniques

Abstract

X-ray computed tomography (CT) is a mainstay in medical imaging, playing major roles in disease diagnosis, injury evaluation, treatment planning, and response assessment. Metal implants, such as a dental filling, artificial hip, spinal fixation rod, or surgical clip, continue to hinder image quality due to their strong attenuation of x-rays, which results in corrupted or missing projection data recorded by the detector. Reconstructing inconsistent or incomplete data due to a metal object yields streaks and defects in the image, known as metal artifacts. Metal artifact reduction (MAR) is one of the remaining important problems in the CT field, especially for radiation and proton therapy planning applications.

This dissertation builds upon work over the past four decades towards reducing CT metal artifacts. We present a novel hybrid approach to improve the acquired data fidelity with an energy-discriminating photon-counting detector (PCD) and perform advanced image-based correction with machine learning algorithms. A PCD measures the energy of every incident x-ray photon, whereas a traditional detector integrates signals received over the entire energy spectrum. By determining individual photon energies, the PCD avoids beam hardening and electronic noise, and also suppresses scatter. A major undertaking of this work has been to upgrade and characterize a benchtop CT system with a PCD for spectral imaging. In parallel, we have developed and optimized deep learning algorithms for mapping between metal-corrupted CT images and artifact-free counterparts. We demonstrate the unique capability of the benchtop system to test and refine algorithms for MAR with physical data. Our methods indicate that higher image quality can be achieved over state-of-the-art MAR methods in challenging cases. Additionally, we have designed and analyzed novel multi-modality imaging techniques for simultaneous acquisition of complementary datasets. These schemes demonstrate the feasibility to overcome single-modality artifacts with perfectly co-registered images.