Assessment of Inter-observer Reproducibility of Radiomic Features in Multiparametric Magnetic Resonance Imaging of Endometrial Cancer

Objective: The aim of our study is to evaluate the inter-observer agreement of radiomic features extracted from different sequences of multiparametric magnetic resonance imaging (mpMRI) in endometrial cancer, thereby contributing to the development of radiomic models with higher reproducibility and facilitating their integration into clinical practice. Materials and Methods: This single-center, retrospective study included 59 patients with pathologically confirmed, untreated endometrial cancer. The MRI scans of the patients were obtained from both in-house and external Picture Archiving and Communication System databases. Three radiologists segmented endometrial cancer tissue on three plane T2 weighted images, sagittal (T2WS), coronal (T2WC) and axial (T2WA), T1 weighted contrast enhanced sagittal (T1CE) images and diffusion weighted (DWI) images with the highest b-value. Apparent diffusion coefficient (ADC) masks were transferred from DWI segmentations. Results: The percentage of radiomic features with excellent agreement (ICC ≥ 0.90) was 58.9% for T2WS, 54.2% for T2WC, 44.9% for T2WA, 57.9% for T1CE, 54.2% for DWI, and 45.8% for ADC. NGTDM and GLCM performed more poorly than the other subclasses, 30% and 45.8% of features being reproducible, respectively. Shape and GLDM performed better than other subclasses, 60.7% and 64.3% of features being reproducible, respectively. Segmentations from T2WS had better reproducibility despite similar levels of Dice Correlation Similarity (DSC) compared to other planes. Conclusion: Among T2-weighted planes, the highest inter-observer agreement was observed in the sagittal plane (T2WS), while the lowest was found in the axial plane (T2WA). Among all sequences, T2WS and T1CE showed the highest agreement, whereas T2WA and ADC demonstrated the lowest. In the subclasses of radiomic features, the highest agreement is observed in Shape and GLDM, while the lowest agreement is seen in NGTDM and GLCM. Overall robustness of MRI derived radiomics features to inter-reader segmentation differences in EC was moderate.