Consensus-based decision tree for prostatic lesion segmentation in radiomics.

Background. Texture analysis allows for increasing the diagnostic accuracy of magnetic resonance imaging (MRI) and improving the differential diagnosis of prostatic lesions. Manual segmentation is commonly used for lesion delineation and subsequent texture analysis. This approach is confounded by high inter-observer variability. Consensus can be used to reduce the variability of prostatic contours. To date, no worldwide standardized consensus-based segmentation algorithms have been proposed. Study goal. To create a consensus-based algorithm for manual prostatic lesion segmentation; to assess the agreement of experts when using the algorithm. Materials and methods. 60 biparametric MRI scans performed per PI-RADS 2.1 and containing PI-RADS 3, 4, and 5 lesions were included in this retrospective study. Two radiologists independently and manually delineated all prostatic lesions using 3D Slicer software. The obtained masks were compared with calculation of the Dice similarity coefficient. For prostatic lesions with DSC greater than or equal to 0.75, the overlap between two masks was selected as the final mask. For prostatic lesions with DSC lower than 0.75, final masks were generated using the developed consensus-based algorithm. Results. The proposed consensus algorithm statistically significantly increased the median DSC value from 0.61 (0.48; 0.73) (p=0.01) to 0.74 (0.62; 0.79) for after algorithm application. Conclusion. The developed decision tree for prostatic lesion segmentation augments existing scientific and methodological literature, allowing objectifying the segmentation process and increasing the DSC value.