Multifaceted analysis of diffusion-weighted magnetic resonance imaging in cervical cancer

Abstract

It is hypothesised that intratumoural heterogeneity in locally advanced cervical cancer (LACC) leads to sections of varying sensitivity to concurrent chemoradiotherapy (CCRT). Therefore, the development of non-invasive and generalizable imaging biomarkers that can assess intratumoural variations of LACC is crucial. Diffusion-weighted imaging (DWI) is a routine sequence included in the Magnetic Resonance Imaging (MRI) assessment of LACC and provides functional information of the tumour microenvironment without the use of exogenous contrast agents. Though the sequence is inherently quantitative, its current clinical use is primarily qualitatively. A deeper analysis of DWI parameters could lead into new insights of intratumoural heterogeneity and potentially improve clinical management of patients with LACC. Firstly, there are multiple compartment models by which diffusion signals can be analysed. A monoexponential model which computes the apparent diffusion coefficient (ADC) is conventionally used. The biexponential IVIM model is thought to provide insights on diffusion and microcirculation perfusion characteristics within a single sequence. However, a major limitation of the IVIM technique is the need to acquire a large set of diffusion-sensitivity weightings, termed b-values, compared to conventional DWI acquisition, which prohibitively increases scan time. In patients with LACC, it was possible to substantially reduce the number of b-values acquired in the IVIM sequence without a significant loss of precision and discriminative ability through simulations and optimizations of in-vivo data. Secondly, radiomic analysis has been proposed as a means of measuring image heterogeneity by calculating an array of imaging features. These features are hypothesised to be reflective of intratumour phenotypic heterogeneity. These inherent image properties offer an unbiased objective assessment of the tumour microenvironment. Associations were observed between DWI radiomic features and various clinicopathological factors important in LACC. Specifically, certain percentiles of the ADC and pure diffusion coefficient (D) demonstrated significant differences between International Federation of Gynaecology and Obstetrics (FIGO) stages and histological sub-types, while certain histogram features of the perfusion fraction (f) were significantly different between tumours with pelvic lymph node (PLN) metastasis and those without. It was also found that ADC, D, and f may be used to track changes in the tumour following CCRT; a lack of change in ADC texture features and f histogram features were associated with poor prognosis. Lastly, radiomic analysis is hypothesized to yield high prognostic power in prediction and classification models. Histological sub-type classification and PLN metastasis predictive models were built using a mixture of DWI and T2W radiomic features and demonstrated moderate performances. However, it was found that some of these features were sensitive to variations in interobserver and intraobserver delineations, so certain features needed to be excluded from further analysis. It was further found that feature values were significantly different between two centres. Despite these differences, it was found that a classification model for histological sub-types built in one centre had similar performance when applied to another centre’s data. In conclusion, DWI and IVIM parameters demonstrate associations with various clinicopathologic factors as well as treatment response. Radiomic analysis of DWI and IVIM parameters produced repeatable features and generalizable classification models.