Click here to


Are you sure ?

Yes, do it No, cancel

Effects of Variability in Radiomics Software Packages in Classifying Patients with Radiation Pneumonitis

J Foy*, J Fuhrman , S Armato , H Al-Hallaq , The University of Chicago, Chicago, IL


(Sunday, 7/29/2018) 2:05 PM - 3:00 PM

Room: Davidson Ballroom B

Purpose: Several open-source radiomics software packages have been developed. This study investigates the effects of using different texture analysis software on the classification of patients with radiation pneumonitis (RP). Knowledge of this variability is crucial for greater harmonization of radiomics research across institutions.

Methods: Pre- and post-radiation therapy (RT) thoracic breath-hold CT scans were collected from 105 esophageal cancer patients. Within each pre-RT scan, 32x32-pixel regions-of-interest (ROIs) were randomly placed to capture normal lung parenchyma that received a mean dose >30Gy. ROIs were placed in anatomically matched locations in the post-RT scans using the vector maps acquired from deformable registration. Six texture features shown to be robust to deformable registration (mean, minimum, GLCM sum average, GLCM sum entropy, GLCM difference entropy, and GLCM entropy) were extracted from each pre- and post-RT ROI using two texture packages: an in-house (“A1�) and an open-source (IBEX) package. Absolute changes in feature values between pre- and post-RT scan ROIs were calculated and averaged across ROIs for each patient. Logistic regression and repeated ANOVA were used to determine relationships between feature value change and the development of RP grade ≥2; significance was assessed at p<0.008 to correct for multiple comparisons. The area under the receiver operating characteristic curve (AUC) was calculated to determine the ability of each feature to classify patients with and without RP.

Results: Individual texture features classified patients with moderate performance for both texture packages (A1: AUC=0.59-0.74; IBEX: AUC=0.58-0.75) with 5 of the 6 features having AUC significantly different from 0.5. AUC values differed significantly between A1 and IBEX for all features.

Conclusion: Significant differences in RP classification ability between two texture packages were shown for all features investigated. While differences in predictive power between packages may not be clinically significant, individual classification tasks may vary based on the software used.


Quantitative Imaging, Texture Analysis, Software


IM/TH- Image Analysis (Single modality or Multi-modality): Quantitative imaging

Contact Email