A Comparative Study of Two Segmentation Methods on Radiomics Feature Robustness in Non-Small Cell Lung Cancer
Vepy Asyana (a*,b), Mohammad Haekal (c) Yeni Pertiwi (a), Deni Hardiansyah (d), Sparisoma Viridi (a), Freddy Haryanto (a), Abdul Waris (a)

a) Nuclear Physics and Biophysics Research Group, Dept. Of Physics, FMIPA ITB, Bandung 40132, Indonesia
*vepyasyana[at]gmail.com
b) Department of Physics, Faculty of Sciences and Mathematics, Universitas Riau, Pekanbaru 28293, Indonesia
c) Department of Physics, ITS, Surabaya 60111 Indonesia
d) Department of Physics, Faculty of Sciences and Mathematics, Universitas Indonesia, Jakarta 40132, Indonesia

Abstract

Radiomics is a mathematical method to convert digital images into mineable data. The data reflects the distribution of signal intensity, texture, or shape of the signal within a specific volume of interest (VOI) or ROI in an image. Analysis of this data can provide information about the biological characteristics of the tumor, patient prognosis, or response to treatment. It has shown exciting promise for improved cancer decision support from early detection to personalized precision treatment and has the potential to revolutionize cancer prognostication. For radiomics, the segmentation method and its effect on radiomic feature stability is a crucial consideration. This study examines the effect of segmentation methods on the stability of radiomic features in identifying differences between Adenocarcinoma and Squamous cell carcinoma cases. This study used PET scan image data of Non-Small Cell Lung Cancer with indications of Adenocarcinoma and Squamous Cell Carcinoma, respectively 25 patients. This image was obtained from The Cancer Imaging Archive (TCIA), a public imaging data source. The segmentation methods used in this study are threshold and edge detection methods. The dice similarity coefficient (DSC) was calculated to quantify segmentation agreement between different segmentation methods. The impact of segmentation on radiomics features was assessed by intra-class correlation coefficients (ICC) and the coefficient of variance (COV). According to the results, the DSC values showed the biggest difference between threshold and edge detection segmentation methods. Grey-level run-length matrix (GLRLM) was a common radiomics signature extracted by all segmentation methods.

Keywords: Radiomics, Non-Small Cell Lung Cancer, PET-Scan

Topic: Biophysics and Medical Physics