Robustness of radiomics features to varying segmentation algorithms in magnetic resonance images

Aim: To verify the accuracy of different segmentation algorithms applied on a dataset of 50 patients suffering from enlargement of the median lobe of the prostate district, to establish whether it is possible to support the work of medical physicians in radiomics analyses through semi-automatic segmentation approaches.Materials and Methods: Seven algorithms were used for prostate segmentation in MR images and for the subsequent extraction of radiomics features. Astatistical analysis was carried out considering the features extracted from semi-automatic and manual segmentations. The analysis was based on the ANOVA test, followed by theTukey test to verify the repeatability of the algorithms, and on the calculation of the intraclass correlation coefficient to verify the reliability and robustness of the extracted features. Based on the correlation between the binary masks extracted for each algorithm and the corresponding binary mask of the medical physicians' segmentation, a volumetric analysis was conducted.Results: The best semi-automatic algorithm to support the medical physician among those evaluated is the "Fill between slices" algorithm, which is also the fastest of all. The least reliable algorithms are those based on the similarity of grey levels.