Manual segmentation methods of biparametric prostate MRI: which should be preferred for texture analysis?

Background. Texture analysis is a rapidly evolving field in prostate cancer diagnosis. Different approaches to prostate segmentation are used for subsequent texture analysis. Semi-automatic and automatic algorithms for prostate lesion segmentation are often inaccessible to radiologists. Therefore, it is necessary to identify an optimal manual segmentation method for fast and reproducible extraction of a maximum number of texture features. Aim. To select an optimal method of manual segmentation for prostate bPMRI for radiomic analysis. Materials and methods. A total of 59 bPMRI were performed per PI-RADS 2.1, containing PI-RADs 3, 4, and 5 lesions, were included in this retrospective study. Two radiologists independently delineated all prostate lesions using contour brush (Draw), round brush (Paint, ROI), and sphere brush (Paint VOI) in 3D slicer. Simultaneously, the segmentation time for each method was measured using Excel table macros, and the median time for each was calculated and compared. Finally, texture features were extracted and compared to ground truth data from our previous study. Results. Median segmentation time using the contour brush was 34.2 (18.67.8) seconds for Radiologist 1 and 31.8 (19.2;55.2) seconds for Radio 2. The median segmentation time for the round brush (ROI) was 27 seconds (16.8;45 seconds) for Radiologist1 and 28 seconds (17;50.8 seconds) Radiologist 2, and for the sphere brush (VOI), it was 16.2 seconds (12.8;20.2 secs) and 15.8 secs (9.8-21.2). Cohen's kappa was less than 0.05. The fastest method of segmentation for both radiologists was the sphere brush. Three second-order features were obtained, which were robust to segmentation variation (p < 0.05). Conclusion. The use of the sphere (VOI) brush greatly simplified and accelerated manual prostate lesion segmentation, compared to the contour brush, while reducing the number of features that are reproducible textures. We identified two features that were robust to variation in segmentation.