Abstract
Objective Identification of anaplastic lymphoma kinase (ALK) and epidermal growth factor receptor
(EGFR) mutation types is of great importance before treatment with tyrosine kinase
inhibitors (TKIs). Radiomics is a new strategy for noninvasively predicting the genetic
status of cancer. We aimed to evaluate the predictive power of 18F-FDG PET/CT-based
radiomic features for mutational status before treatment in non-small cell lung cancer
(NSCLC) and to develop a predictive model based on radiomic features.
Methods Images of patients who underwent 18F-FDG PET/CT for initial staging with the diagnosis
of NSCLC between January 2015 and July 2020 were evaluated using LIFEx software. The
region of interest (ROI) of the primary tumor was established and volumetric and textural
features were obtained. Clinical data and radiomic data were evaluated with machine
learning (ML) algorithms to create a model.
Results For EGFR mutation prediction, the most successful machine learning algorithm obtained
with GLZLM_GLNU and clinical data was Naive Bayes (AUC: 0.751, MCC: 0.347, acc: 71.4%).
For ALK rearrangement prediction, the most successful machine learning algorithm obtained
with GLCM_correlation, GLZLM_LZHGE and clinical data was evaluated as Naive Bayes
(AUC: 0.682, MCC: 0.221, acc: 77.4%).
Conclusions In our study, we created prediction models based on radiomic analysis of 18F-FDG
PET/CT images. Tissue analysis with ML algorithms are non-invasive methods for predicting
ALK rearrangement and EGFR mutation status in NSCLC, which may be useful for targeted
therapy selection in a clinical setting.
Keywords 18F-FDG-PET/CT - Lung cancer - Textural analysis - EGFR - ALK
