Impact of inter-observer variability on MRI radiomic features in bladder cancer

Abstract

Purpose or Objective Muscle-invasive bladder cancer (MIBC) can be radically treated with either surgery or radiotherapy. Regardless of treatment, overall 5-year survival remains poor with 1 in 2 patients dying of disease. Improved patient stratification is needed, identifying those who respond better to radiotherapy. Using imaging biomarkers from on-treatment MRI such as radiomic features is one viable strategy. We aimed to identify MRI radiomic features robust for inter-observer variation in MIBC patients treated with chemo-radiotherapy and compare bladder wall (BW) and tumour bed (TB). Materials and Methods Ten eligible MIBC patients were recruited to an MRI imaging study while undergoing chemo-radiotherapy. All patients were treated with GemX (52.5 /20 fractions with Gemcitabine 100 mg/m2 weekly). Weekly scans were taken during treatment on a Siemens 1.5T diagnostic MRI scanner. BW and TB were segmented as the region of interests on the weekly T2-weighted MRI images by two clinical oncologists to evaluate inter-observer variability. In total, 40 MRI images were available to extract features using pyRadiomics (v3.0.1). Robustness to inter-observer variation was assessed using the intraclass correlation coefficient (ICC). Distribution of features in BW and TB was compared using inter-quartile range.