Open AccessPET Radiomic Features Variability: A Phantom Study on the Influence of Reconstruction and Discretization Method
Author(s) -
Marcos Machado,
Jean Lucas Pereira Pita,
Eduardo Martins Netto
Publication year - 2021
Publication title -
revista brasileira de física médica
Language(s) - English
Resource type - Journals
eISSN - 2176-8978
pISSN - 1984-9001
DOI - 10.29384/rbfm.2021.v15.19849001598
Subject(s) - discretization , imaging phantom , bin , entropy (arrow of time) , computer science , artificial intelligence , iterative reconstruction , pattern recognition (psychology) , mathematics , nuclear medicine , algorithm , physics , medicine , mathematical analysis , quantum mechanics
We aimed to explore the variability of PET radiomic features for varying reconstruction methods and quantification settings. The IQ-NEMA phantom was scanned 5 times with a sphere to background F-18 concentration ratio of 10:1. The activity and the image duration were matched to result typical counting statistics for 18F-FDG oncologic examinations. The images were reconstructed with OSEM and PSF reconstructions, then 99 radiomic features were extracted using two discretization methods: fixed bin number (FBN = 16, 32 and 64 gray levels) and fixed bin width (FBW=0.25). This scheme resulted in a total of 1,188 features, classified as having low (<5.0%), intermediate (5-29.9%) or high (≥30%) variability. In general, FBW discretization yielded more stable features. A total of 499, 558 and 131 features had low, intermediate and high variability, respectively. First order features such as energy and entropy and textural features such as entropy (GLCM), long run emphasis and short run emphasis (GLRLM) were more likely to present low variability, regardless the reconstruction and discretization method. Other textural features such as large area emphasis (GLSZM), zone percentage (GLSZM) and complexity (NGTDM) had more frequently intermediate or high variability.These findings could facilitate features’ selection for further PET radiomic applications.