Supervisors info:
Αναστάσιος Γαϊτάνης, Ειδικός Λειτουργικός Επιστήμονας Β', Κέντρο Κλινικής Έρευνας, Πειραματικής Χειρουργικής & Μεταφραστικής Έρευνας, Ίδρυμα Ιατροβιολογικών Ερευνών Ακαδημίας Αθηνών
Ευστάθιος Ευσταθόπουλος, Καθηγητής, Ιατρική Σχολή, Εθνικό και Καποδιστριακό Πανεπιστήμιο Αθηνών
Ελλάς Σπυράτου, Μεταδιδακτορική Ερευνήτρια, Ιατρική Σχολή, Εθνικό και Καποδιστριακό Πανεπιστήμιο Αθηνών
Summary:
Radiomics is a newly emerging field, claiming that there is a lot more information than a “naked” eye can conceive in medical images and needs to be extracted in the form of texture features. Medical images acquired through any imaging system (e.g. Positron Emission Tomography (PET), Computed Tomography (CT), Ultrasound (US), or Magnetic Resonance (MR)) are suitable candidates. However, a lot of research is required to identify the texture indices, which firstly, under certain conditions are reliable and secondly, can be associated with certain biological processes. Radiomics aim to find those indices and use them in diagnosis, prognosis, and treatment evaluation (theranostics). In this research work, we shall investigate, under preclinical conditions, the impact of bin size (total number of grey levels) in texture features and the repeatability, reproducibility, and liability of them (under specific conditions, both in terms of image reconstruction and acquisition time) using the NEMA NU 4-2008 image quality (IQ) phantom. For the acquisition of the PET images a small-animal PET/CT scanner – Mediso nanoScan®PC (PET8/2) – was employed. LIFEx version 6.20 (an open-source medical imaging analysis software) was incorporated for the analysis of PET images and the extraction of texture features (First-Order features, Second-Order features, and Higher-Order features). Statistical analysis (PD, COV, Interclass Correlation Coefficient, Spearman Rank Correlation Coefficient, one-way Analysis of Variance, and Least Significance Test) and evaluation of the indices was performed. After the evaluation with IQ phantom images, a preliminary study using PET image data from treated and non-treated mice was performed. The vast majority of the texture features showed low-to-no robustness and a high degree of variance, thus, making them bad candidates for potential biomarkers. Only a limited number of features found to perform well at all statistical tests (Histogram Skewness, Kurtosis, and Excess Kurtosis, GLCM Correlation, and NGLDM Coarseness), while, even less (Histogram Skewness and GLCM Correlation) could provide valuable information to distinguish the mice of the Treated from the Non-Treated group. Furthermore, we are proposing that the robustness of the features is directly associated with the total number of bins – First-Order and Second-Order features tend to be more robust at a higher number of bins, while Higher-Order features show greater robustness at a lower number of bins.
Keywords:
Radiomics, Texture analysis, Image analysis, Small-animal PET imaging, 18F-FDG, Texture indices, Texture features, Cancer, Biomarkers, LIFEx, NEMA NU 4-2008 image quality (IQ) phantom
