Michele Larobina1, Rosario Megna2, Raffaele Solla1. 1. Institute of Biostructures and Bioimaging, National Research Council (CNR), Via Tommaso de Amicis, 95, 80145, Napoli, Italy. 2. Institute of Biostructures and Bioimaging, National Research Council (CNR), Via Tommaso de Amicis, 95, 80145, Napoli, Italy. rosario.megna@cnr.it.
Abstract
PURPOSE: To compare texture feature estimates obtained from 18F-FDG-PET images using three different software packages. METHODS: PET images from 15 patients with head and neck cancer were processed with three different freeware software: CGITA, LIFEx, and Metavol. For each lesion, 38 texture features were extracted from each software package. To evaluate the statistical agreement among the features across packages a non-parametric Kruskal-Wallis test was used. Differences in the features between each couple of software were assessed using a subsequent Dunn test. Correlation between texture features was evaluated via the Spearman coefficient. RESULTS: Twenty-three of 38 features showed a significant agreement across the three software (P < 0.05). The agreement was better between LIFEx vs. Metavol (36 of 38) and worse between CGITA and Metavol (24 of 38), and CGITA vs. LIFEx (23 of 38). All features resulted correlated (ρ > = 0.70, P < 0.001) in comparing LIFEx vs. Metavol. Seven of 38 features were found not in agreement and slightly or not correlated (ρ < 0.70, P < 0.001) in comparing CGITA vs. LIFEx, and CGITA vs. Metavol. CONCLUSION: Some texture discrepancies across software packages exist. Our findings reinforce the need to continue the standardization process, and to succeed in building a reference dataset to be used for comparisons.
PURPOSE: To compare texture feature estimates obtained from 18F-FDG-PET images using three different software packages. METHODS: PET images from 15 patients with head and neck cancer were processed with three different freeware software: CGITA, LIFEx, and Metavol. For each lesion, 38 texture features were extracted from each software package. To evaluate the statistical agreement among the features across packages a non-parametric Kruskal-Wallis test was used. Differences in the features between each couple of software were assessed using a subsequent Dunn test. Correlation between texture features was evaluated via the Spearman coefficient. RESULTS: Twenty-three of 38 features showed a significant agreement across the three software (P < 0.05). The agreement was better between LIFEx vs. Metavol (36 of 38) and worse between CGITA and Metavol (24 of 38), and CGITA vs. LIFEx (23 of 38). All features resulted correlated (ρ > = 0.70, P < 0.001) in comparing LIFEx vs. Metavol. Seven of 38 features were found not in agreement and slightly or not correlated (ρ < 0.70, P < 0.001) in comparing CGITA vs. LIFEx, and CGITA vs. Metavol. CONCLUSION: Some texture discrepancies across software packages exist. Our findings reinforce the need to continue the standardization process, and to succeed in building a reference dataset to be used for comparisons.