MRI-based radiomics: Quantifying the stability and reproducibility of tumour heterogeneity in vivo and in a 3D printed phantom

Abstract

Magnetic resonance imaging (MRI) is a key component in the oncology workflow. Radiomics analysis is a new approach that uses standard of care (SOC) magnetic resonance (MR) images to non-invasively characterise tumour heterogeneity. For radiomics to be reliable, the imaging features measured must be stable and reproducible. This thesis aims to quantify the stability and reproducibility of MRI-based radiomics in vivo and in a 3D printed phantom. Chapter 4 explores the feasibility of constructing a 3D printed phantom using an MRI visible material (‘red resin’). The study shows that the material used to construct an anthropomorphic skull phantom mimicked human cortical bone with a T2* of 411 ± 19 µs. The phantom material provided sufficient signal for tissue segmentation however was only visible with an ultrashort echo time sequence, not commonly used in SOC imaging. Chapter 5 investigates a high temperature resin (‘white resin’) where a texture object was developed for analysis. The ‘white resin’ was visible using SOC sequences. The interscanner repeatability measurements of the texture phantom demonstrated high reproducibility with 76% of texture features having an ICC > 0.9. In chapter 6, further texture and shape objects were developed and employed in a multi-centre study assessing inter and intrascanner variation of MRI-based radiomics. The phantom was stable over a period of 12 months, with a T1 and T2 of 150.7 ± 6.7 ms and 56.1 ± 3.9 ms, respectively. The study also found that histogram features were more stable (ICC > 0.8 for 67%) compared to texture (ICC > 0.8 for 58%) and shape texture (ICC > 0.8 for 0%) across the 8 scanners. In chapter 7, phantom measurements found that radiomics features were more sensitive to changes of image resolution and noise. The in vivo test-retest component of chapter 7 detected many unstable features not suitable for use in a radiomics prognostic model. In chapter 8, of the 83 features computed only 19 features had significant changes between the baseline, mid and post radiation treatment and may be informative to assess rectal cancer treatment response. When considering using radiomics analysis for SOC MRI scans, caution must be taken to ensure imaging protocols, imaging equipment including scanners and coils are consistent to improve intra and inter-institutional feature robustness. This can be achieved with regular quality assurance (QA) of imaging protocols using a suitable phantom and appropriate feature selection using phantom and in vivo datasets.