Objective:

To identify epidermal growth factor receptor (

EGFR

) mutation status lung adenocarcinoma patients using

18

F-FDG positron emission to

mography (PET)/computed tomography (CT) images from two different institutions

and to find appropriate imaging preprocessing methods to handle CT images of different slice thickness.

Methods:

18

F-FDG PET/CT images of 148 patients were retrospectively studied

which were scanned in two hospitals with different scanners and CT slice thickness (3 and 5 mm

respectively). The tumor region of interest (ROI) was manually delineated. Gray scale unifying

downsampling and upsampling with interpolation were used to pre-processing CT images. After the extraction and selection of radiomic features

logistic regression (LR) models were constructed to predict

EGFR

mutation status. Feature stability was evaluated through statistical tests. LR models were evaluated using the receiver operating characteristic (ROC) curve and served as an indicator of feature predicting power.

Results:

All adopted methods improves feature stability. The combination of gray scale unifying and downsampling performed best

resulting in 30.8% more robust features compared to baseline

and the corresponding LR model achieves area under curve (AUC) scores 0.862 in training and 0.716 in test cohort.

Conclusion:

The combination of gray scale unifying and downsampling is an effective approach for multi-center radiomic analysis

improving both feature stability and predictive power.