Room: Exhibit Hall | Forum 8
Purpose: The objective of this study was to explore radiomics features for the prediction of radiation pneumonitis (RP) development in non-small cell lung cancer (NSCLC).
Methods: We prospectively enrolled 126 cases with NSCLC into a cohort study which underwent computed tomography (CT) before CRT. For each patient, diagnostic CT scans were acquired before RT and during RT. Two groups of tumor features were examined: (1) clinical features (eg, TNM stage, age and gender) and demographics; (2) spatial texture features of CT, which characterize tumor intensity range, spatial patterns and distribution and associated changes resulting from CRT. A reproducible and no redundant feature set was statistically filtered and validations. The machine learning modeling were used to test the relationships between radiomics features and development of grade >=2 RP. Receiver operating characteristic (ROC) curve and the area under curve (AUC) analysis were performed to determine the overall performance of extracted texture features in whether radiomics feature combinations could improve RP distinction.
Results: The models analyses showed that the radiation dose; IHIST_energy, m_contrast_2, m_entropy_2, Diff_homogeneity_2, m_lnversevariance_2, high intensity small zone emphasis (HISE) and low intensity large zone emphasis (LILE) were associated significantly with the RP. Radiomics features could discriminate between patients with and those without RP (AUCs from 0.51 to 0.74). Using artificial neural networks in a classifier model, AUC has a higher increased (0.62-0.86).
Conclusion: The heterogeneity of normal lung tissues assessed by texture analysis of pre-treatment CT, may has the potential to act as a novel predictor of â‰¥2 radiation pneumonitis for NSCLC treated by radiotherapy.
Not Applicable / None Entered.