Abstract:
Nasopharyngeal cancer (NPC) is a type of head and neck cancer that is prevalent in China and Southeast Asia. The standard treatment for NPC is radiation therapy (RT), which requires accurate target volume delineation (TVD) to ensure an effective radiation dose while minimizing damage to surrounding healthy tissues. However, TVD is subject to intra- and inter-observer variability, which can affect the accuracy of RT and lead to unstable results in radiomics analysis. Radiomics is a technique that extracts quantitative data from medical images, but it is currently limited by the manual process of tumor delineation. This study aims to investigate the impact of inter-observer variability in manual tumor delineation on the stability of radiomic features in NPC. CT images were collected from patients who were diagnosed with Nasopharyngeal carcinoma (NPC) and were manually contoured by radiation oncologists to determine gross tumor volume (GTV) of the primary tumor. CT images were then analyzed to extract radiomic features using Pyradiomics. The extracted features were classified into four feature classes, and the binwidth parameter was varied into three values. To measure inter-observer variability, Dice similarity coefficient, intraclass correlation coefficient (ICC), and coefficient of variation (CV) were used. The ICC threshold of 0.8 was selected to determine stable features, and the CV threshold of 10% was selected for excellent reproducibility. The results showed that the shape and first-order classes were stable in all binwidths with an ICC value greater than 0.8, indicating good performance in tumor delineation. Additionally, the wavelet class had the most robust feature with respect to ICC and %CV, which can be used for prediction models in the future.
