Texture analysis of multiparametric MRI and association with tumour grading in cervical cancer

Abstract

Purpose: To explore efficacy and generalizability of texture analysis of multiparametric magnetic resonance imaging (mpMRI) in discriminating histopathological characteristics of cervical cancers in a two-centre setting. Methods and Materials: 130 cervical cancer patients were retrospectively reviewed for pre-treatment diffusion-weighted (DWI) and standard T2-weighted (T2W) abdominopelvic MRI in two centres (centre 1 n=100; centre 2 n=30). Biopsies of each patient were acquired to determine histological subtype and tumour grading. Volumetric regions of interest (VOI) were placed to encompass the entirety of primary tumours on T2W and DWI parametric maps. 144 texture features were calculated using pyradiomics. Redundancy analysis was used for feature reduction while goodness-of-fit measures were used for feature selection. Logistic regression was used to build predictive models with centre 1 serving as the training set and centre 2 serving as an external testing set. Results: In the first centre 80 patients had squamous cell carcinoma (SCC) while 20 had adenocarcinoma (ACA), and 42 patients had well or moderately differentiated tumours (G1/G2) while 58 had poorly differentiated tumours (G3). In the second centre, 25 patients had SCC while 5 had ACA, and 14 had G1/G2 tumours while 16 had G3 tumours. Redundancy analysis demonstrated that only 15% of features were independent predictors. In discriminating histological grading, only one texture, T2Wkurtosis, achieved reasonable goodness-of-fit and moderate discriminative performance (internal accuracy: 0.64, external accuracy: 0.71). In discriminating histological subtype. Texture features had poor discriminative performance for tumour grading. Conclusion: T2Wkurtosis was moderately discriminative in differentiating tumour grading in cervical cancer.