Study Subjects
All patients included in this study were diagnosed with NPC and underwent FDG PET/CT scans before and after definitive radiotherapy (with or without chemotherapy), between 2008 and 2019. The study population also met the following criteria: (1) subjects were more than 20 years of age; (2) the interval between definitive radiotherapy and follow-up FDG PET/CT was less than 6 months; and (3) the PET images were acquired with the same spatial resolution within the study population. We excluded patients who were treated for recurrence before the follow-up FDG PET/CT. Patients with the PET/CT image data loss were also excluded. Consequently, 145 patients were included in this study.
This study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of Samsung Medical Center (protocol code 2020-06-127; date of approval: 15 July 2020). The requirement of written informed consent from enrolled subjects was waived by the Institutional Review Board of Samsung Medical Center due to the retrospective study design.
Medical Record Review
In this retrospective study, medical records were reviewed for demographic/clinical characteristics and information on recurrence/death among NPC patients. Follow-up of records review included the 60 months after definitive radiotherapy. Progression-free-survival (PFS) was defined as the duration of definitive radiotherapy completion and recurrence in months. Date of recurrence was determined as the date of documented clinical decision based on imaging modalities. Overall survival (OS) was defined as the duration to definitive radiotherapy completion and death from all causes in months. Cancer staging was based on the Sixth Edition of the American Joint Committee on Cancer’s Cancer Staging Manual to 2009, the Seventh Edition from 2010 to 2017, and Eighth Edition after 2018.
FDG PET/CT Imaging
Patients fasted for at least 6 hours before the injection of FDG, and their blood glucose concentration was confirmed to be less than 200 mg/dL at the time of FDG injection. At 60 minutes after injection of 5 MBq/kg of FDG, imaging was performed using an STe PET/CT scanner (GE Healthcare) without intravenous or oral contrast. Whole-body CT images were obtained with a continuous spiral 16-slice helical CT technique (140 keV; 30–170 mA; section width, 3.75 mm). An emission PET scan was obtained from the level of the thigh to the skull base. Scanning was performed at 2.5 min per frame in 3-D mode with attenuation-corrected images (3.9 x 3.9 x 3.3 mm) reconstructed using a 3-D ordered-subset expectation maximization algorithm (20 subsets, 2 iterations).
Imaging Analyses
Primary tumor segmentation of all pre- and post-treatment FDG PET images was computed using the gradient-based segmentation method (‘PET edge’) of MIM version 6.4 (MIM software, Inc., Cleveland, OH, USA). Only primary tumors were analyzed in this study. A total of 72 quantitative radiomic features were extracted from the PET images using Chang-Gung Image Texture Analysis (CGITA) toolbox [12]. The relative variation of each radiomic feature was also evaluated in pre- and post-treatment scans:
$${\Delta } \text{f}\text{e}\text{a}\text{t}\text{u}\text{r}\text{e}=\frac{{feature}_{post-treatment}-{feature}_{pre-treatment}}{{feature}_{pre-treatment}}$$
Statistical Analyses
The study population was divided randomly into two groups, the training and test sets (7:3) using the createDataPartition function of caret package in R. The demographic and clinical characteristics were compared between two groups using Student’s t-test, Fisher’s exact test, the chi-square test, and Wilcoxon rank sum test. Two-sided p values less than 0.05 were considered significant. A random survival forest (RSF) model was adopted to perform survival analyses of PFS and OS in the training set. An RSF is a nonparametric method to analyze right-censored survival data with multiple covariates and there is no need to fulfill any assumptions such as the proportional hazards assumption of classical Cox regression [13]. An RSF provides a variable importance (VIMP) that demonstrates variables of key role in predicting the survival outcome. First, we developed an RSF model with clinical variables—age, sex, smoking history, p16 status, EBV status, and stage— and conventional PET parameter—maximum SUV, mean SUV, MTV, and TLG—from pre- and post-treatment scans. Next, we made another RSF model with clinical variables and selected 10 relatively important radiomic PET features. Each model was then evaluated in the test set. Brier score (BS; mean squared error) was used to assess the prediction performance of the models in the training and test sets. Time-dependent BS and continuous ranked probability scores (CRPS; integrated BS divided by time) were recorded.
All statistical analyses were performed using R version 4.1.2 (The R Foundation for Statistical Computing, Vienna, Austria), with caret version 6.0.92, dplyr version 1.0.9, psych version 2.2.5, tidyverse version 1.3.1, gmodels version 2.18.1.1., survival version 3.3.1, randomForestSRC version 3.1.0 packages.