Encrypted login | home

Program Information

Reproducibility and Robustness of Radiomic Features Extracted with Semi-Automatic Segmentation Tools

C Owens

C Owens1,2*, C Tang1 , C Peterson1 , X Fave1,2 , E Koay1 , M Salehpour1 , D Fuentes1 , J Li1 , L Court1 , J Yang1 , (1) University of Texas MD Anderson Cancer Center, Houston, TX, (2) UT Health Science Center Graduate School of Biomedical Sciences, Houston, Texas


WE-F-205-11 (Wednesday, August 2, 2017) 1:45 PM - 3:45 PM Room: 205

Purpose: To evaluate the inter-observer variability and software dependency of radiomic features for lung tumors extracted from CT images using both manual segmentation and semi-automatic segmentation.

Methods: Three radiation oncologists manually delineated lung tumors from CT images of 10 patients using two different software tools (3D-Slicer and MIM Maestro). In addition, two observers without formal clinical training were instructed to use two semi-automatic tools, Lesion Sizing Toolkit (LSTK) and GrowCut in 3D-Slicer, to delineate the same tumor volumes. Thirty-one radiomics features were calculated for the tumor contours using IBEX software. The impact of inter-observer and inter-software variability on feature reproducibility was evaluated using intra-class correlation coefficients (ICC). Feature robustness was evaluated using the feature range.

Results: Radiomic features extracted using semi-automatic segmentations had higher inter-observer reproducibility (ICC₌0.987±0.012) compared to features extracted from manual segmentations (ICC₌0.889±0.145). For inter-software variability, the average ICC value for each observer was within good reproducibility bounds (ICC≥0.60), whereas for the manual method the average ICC value varied greatly for each observer. For semi-automatic contours, 87% of features had smaller feature ranges, respectively, compared to those extracted from manual contours.

Conclusion: Radiomic features extracted using semi-automatic segmentation are less variable and more robust among observers. With semi-automatic segmentation tools, observers without formal clinical training were comparable to physicians in tumor delineation.

Funding Support, Disclosures, and Conflict of Interest: This work was partially support by the CPRIT grant RP110562-P2.

Contact Email: