Identification Of Radiomic Biomarkers For Patients With Locally Advanced Lung Cancer

Radiomic analysis of the pre-treatment CT scans of lung cancer patients has the potential to provide valuable insight into patient outcomes. We hypothesized that there would be an association between pre-treatment radiomic features and overall survival following chemoradiation for locally advanced lung cancer. Twenty-four patients with stage III lung cancer (21 non-small cell, 3 small cell) who received cisplatin/etoposide and concurrent thoracic radiotherapy to a mean dose of 66 Gy (range 58-74 Gy) with accelerated fractionation (6 fractions/week) were treated as part of an IRB-approved study. The gross tumor volume (GTV) was contoured on pre-treatment free-breathing CT images, from which 61 quantitative radiomic features were extracted to collectively capture tumor morphology, intensity, fine-texture, and coarse-texture. All radiomics features were standardized using Z-score normalization. Principal component analysis (PCA) was used to characterize the inter-feature correlation. Leave-one-out cross-validation and least absolute shrinkage and selection operator (LASSO) regularized Cox regression were performed to study the multivariable relationship between radiomics features and overall survival. After a median follow-up of 57 months in living patients, 17 of 23 patients died. Following PCA, 4 radiomic features were found to represent 85% of the variance in the radiomics data. Leave-one-out cross-validation identified 11 features that appeared in ≥1 of the 24 LASSO models, and five features (LongRunHighGrayLevelEmphasis, SmallSizeLowGrayLevelEmphasis, VariationOfIntensity, Entropy, and Contrast) appeared in ≥15 of these models. Stepwise Cox proportional hazards models fit with the data and 11 pre-selected features in the leave-one-out process. According to Akaike information criterion (AIC), the following four radiomic features in the final model were associated with overall survival: LongRunHighGrayLevelEmphasis (HR 3.30, 95%CI 1.53-7.15, p=0.0024), SmallSizeLowGrayLevelEmphasis (HR 0.19, 95%CI 0.06-0.63, p=0.0061), VariationOfIntensity (HR 6.19, 95%CI 1.85-25.84, p=0.0041), and GrayLevelNonUniformity (HR 0.41, 95%CI 0.17-1.02, p=0.0557). Overall, these results suggest that relatively dense tumors with a homogenous coarse texture were associated with worse overall survival. The pre-chemoradiation planning CT-based radiomic features used in this study were found to be associated with overall survival in patients with stage III lung cancer. This hypothesis generating study was uniquely bolstered by prospective clinical trial design and mature clinical follow-up, but was limited by low overall patient numbers. Our work provides the groundwork for trials in larger patient populations and reinforces the utility of radiomics as a potentially powerful prognostic tool.