Predicting programmed death-ligand 1 expression level in non-small cell lung cancer using a combination of peritumoral and intratumoral radiomic features on computed tomography

Abstract

In this study, we investigated the possibility of predicting expression levels of programmed death-ligand 1 (PD-L1) using radiomic features of intratumoral and peritumoral tumors on computed tomography (CT) images. We retrospectively analyzed 161 patients with non-small cell lung cancer. We extracted radiomics features for intratumoral and peritumoral regions on CT images. The null importance, least absolute shrinkage, and selection operator model were used to select the optimized feature subset to build the prediction models for the PD-L1 expression level. LightGBM with five-fold cross-validation was used to construct the prediction model and evaluate the receiver operating characteristics. The corresponding area under the curve (AUC) was calculated for the training and testing cohorts. The proportion of ambiguously clustered pairs was calculated based on consensus clustering to evaluate the validity of the selected features. In addition, Radscore was calculated for the training and test cohorts. For expression level of PD-L1 above 1%, prediction models that included radiomic features from the intratumoral region and a combination of radiomic features from intratumoral and peritumoral regions yielded an AUC of 0.83 and 0.87 and 0.64 and 0.74 in the training and test cohorts, respectively. In contrast, the models above 50% prediction yielded an AUC of 0.80, 0.97, and 0.74, 0.83, respectively. The selected features were divided into two subgroups based on PD-L1 expression levels ≥ 50% or ≥ 1%. Radscore was statistically higher for subgroup one than subgroup two when radiomic features for intratumoral and peritumoral regions were combined. We constructed a predictive model for PD-L1 expression level using CT images. The model using a combination of intratumoral and peritumoral radiomic features had a higher accuracy than the model with only intratumoral radiomic features.

Investigation of the combination of intratumoral and peritumoral radiomic signatures for predicting epidermal growth factor receptor mutation in lung adenocarcinoma

“…These segmentations were confirmed by a radiation oncologist with over 16 years of experience in radiation therapy. The peritumoral region was determined using quantitative morphologic operations as a radially extending region with 3, 5, and 7 mm radius from the intratumoral region of the tumor boundary 19 . These peritumoral regions included air in the lungs, pulmonary vessels, and bronchi and did not include the thoracic wall and mediastinum.…”

“…The major settings for radiomic features extraction were as follows: bin width of feature extraction parameters, 30 22 ; sigma size for the LoG filter, 1.0 or 3.0 mm; bin width of the wavelet filter, 10. ResamplePixelSpacing was set to none 19 . In total, 1046 radiomic features were extracted from intratumoral and peritumoral regions using the above conditions.…”

“…ResamplePixelSpacing was set to none. 19 In total, 1046 radiomic features were extracted from intratumoral and peritumoral regions using the above conditions. The combination of intratumoral and peritumoral regions included 2092 radiomic features.…”

“…was determined using quantitative morphologic operations as a radially extending region with 3, 5, and 7 mm radius from the intratumoral region of the tumor boundary. 19 These peritumoral regions included air in the lungs, pulmonary vessels, and bronchi and did not include the thoracic wall and mediastinum. Radiomic features were extracted from intratumoral and peritumoral regions using the open-source software Pyradiomics (version 3.7.1) in Python.…”