Association of robust radiomic features from staging¹⁸F-FDG PET/CT in lung cancer with EGFR expression and overall survival

Purpose: Extraction of stable radiomic features (RFs) from ¹⁸F-FDG PET/CT is challenging due to different acquisition, processing and post-processing protocols. In this report, we aimed to evaluate robustness of RFs, extracted using Image Biomarker Standardization Initiative (IBSI)-compliant approach, to 4 different primary tumor segmentation methods, in patients with lung cancer and to assess any association of TFs with specific molecular marker expression and overall survival (OS).

Methods: Under IRB approval, 87 patients (18 females, 69 males; age 60±10 yrs, range: 30-81 yrs) with treatment naïve primary lung cancer were included in this prospective study. All patients underwent a pretherapy ¹⁸F-FDG PET/CT using a standardized clinical protocol.Primary tumors were delineated using 4 different segmentation methods: Method 1, manual; Method 2: manual with peripheral 1 mm erosion; Method 3: absolute threshold at SUV 2.5; Method 4: relative threshold at 40% SUVmax.Radiomic features (RFs) were extracted using LifeX v.5.1. PET images were spatially resampled into 3 mm isotropic voxels and the grey levels within the extracted volume of interest (VOI) were normalized between the minimum and maximum SUV values and discretized into 64 bins. A total of 35 RFs per segmentation method were evaluated.Before testing for TF reproducibility, the study cohort was randomly divided into 2 groups (testing & validation) in a ratio of 1:2 (n = 29 & n=58, respectively). Testing cohort was used to identify robust RFs, defined as having a minimum concordance correlation coefficient (CCC) ≥ 0.75 among all the 4-segmetnation methods. The resulting TFs were used for subsequent analysis in the validation cohort (n=58). Association of these quantitative TFs with EGFR expression and other histopathologic characteristics was tested using Mann-Whitney U test. All patients in the validation cohort had a follow-up of 2-years since the date of PET/CT scan or till death. Cox Proportional Hazard model was used to evaluate the association of different RFs with 2-year overall survival (OS).

Results: Testing and validation cohorts were equivalent regarding basic characteristics (age, sex & tumor stage).Twelve RFs were deemed robust to the described 4 segmentation methods: SUV SD, SUVmax, SUVpeak, TLG, histogram energy uniformity, volume, compacity, GLCM energy, GLRLM GLNU, GLRLM RLNU, NGLDM busyness, and GLZLM LZHGE.Shape features (tumor volume and compacity) from any segmentation method were associated with EGFR expression.At the end of 24-months of follow-up, 41 patients died and 17 were still alive (OS=29.3%; median survival = 14.7 months, 95%CI: 10.1-19.3 months). Five TFs were associated with 2-year OS from any of the 4 segmentation methods: TLG, volume, compacity, GLRLM GLNU and GLRLM RLNU. The highest hazard ratio for OS was consistently generated from tumor volume and compacity regardless the segmentation method.

Conclusions: Segmentation method of the primary lung tumors on pretherapy ¹⁸F-FDG PET/CT can affect radiomic feature stability. These preliminary results showed that only 12 features were robust enough to 4 different segmentation approaches; two of these features (tumor volume and compacity, which reflects how compact the tumor volume is) showed significant association with EGFR expression and overall survival in patients with lung cancer.Future work will further validate these findings in larger population, add CT features and employ machine learning models to optimize prediction of molecular marker expression and prognosis in terms of response and survival.

• Download figure
• Open in new tab
• Download powerpoint