Multiparametric MRI assessment of cervical carcinoma

Abstract

Diffusion-weighted imaging (DWI) serves as a pillar of magnetic resonance imaging (MRI) diagnosis and evaluation which permits detection of subtle changes at the cellular level, particularly aids in oncological imaging. Quantitative imaging parameters such as multiple diffusion coefficients, T1 and T2 values generated from multiparametric MRI play critical role in the assessment of cervical carcinoma (CC). The series of original studies included in this thesis aimed to investigate and assess the clinical value of multiparametric MRI incorporated with radiomics analysis in the clinicopathological characterisation of CC, as well as to explore the feasibility of magnetic resonance fingerprinting (MRF) technique in tissues differentiation between CC and normal cervix.

Firstly, as an extension of conventional DWI, diffusion kurtosis imaging (DKI) has been a promising model to capture the non-Gaussian movement of water molecules in biological tissues. Preliminary study based on DKI model (n = 117) was conducted to assess the ability of DKI in the differentiation of clinicopathological characteristics of CC, including histological subtype, tumour grade and International Federation of Gynecology and Obstetrics (FIGO) stage. We found that mean apparent diffusion coefficient (ADC), mean diffusivity (MD) and mean kurtosis (MK) were significantly different between histological subtypes. Moreover, mean ADC and MD could differentiate tumour grades, while MK could not.

Secondly, radiomics is a quantitative analysis involving extraction, analysis and modelling of a variety of features derived from medical images related to prediction targets. Radiomics opens up an environment ideal for machine learning and data-based research. Texture analysis (n = 95) and radiomics study (n = 117) based on multiparametric MRI were investigated in CC. The results showed that first-order texture features extracted multiple MRI sequences, including T2-weighted imaging (T2WI), ADC and contrast-enhanced T1-weighted (T1c) imaging, were useful in discriminating clinicopathological characteristics of CC, with excellent performance of the support vector machine (SVM) models. In addition, combined radiomic features extracted from both T2WI and DKI using random forest (RF) model exhibited excellent diagnostic efficiency for histological subtyping. However, DKI-only RF model demonstrated the highest area under the curve (AUC) for FIGO staging.

Lastly, MRF is a fast and flexible framework which yields the estimation of multiple tissue properties within one single and time-efficient scan. Feasibility MRF study in normal cervix and CC was performed (n = 12 and 28, respectively). T1 and T2 estimates generated from MRF exhibited excellent scan-rescan repeatability in normal cervical tissues. Furthermore, we observed significant differences in T1 and T2 estimates between normal cervical tissues and CC, suggesting the potential of MRF in distinguishing CC from normal cervical tissues.

In summary, multiparametric MRI opens up the opportunity for quantitative analysis that reduces subjectivity in aiding the comprehensive assessment of CC. Encouraging findings were achieved in multiparametric MRI and relevant radiomics-based studies for the clinicopathological characterisation of CC.