Session Item

To develop radiomics models for predicting brain metastases (BM) development in patients with stage III non-small cell lung cancer (NSCLC) using the planning contrast-enhanced chest CT for thoracic radiotherapy (TRT).

Patients with stage III NSCLC treated in Maastro between 2012-2021 were screened. Eligibility criteria: adequately staged with 18FDG-PET and brain imaging, treated with definitive chemoradiotherapy, no thoracic surgery before TRT. Exclusion criteria: other malignancy within 5 years; no available planning CT. The time to BM was calculated from pathological diagnosis to imaging confirmed BM or last follow-up if no event was observed. The regions of interest (ROIs) were the gross tumor volume (GTV), primary lung tumor (GTVp), and metastatic lymph nodes (GTVn). GTV was the sum of GTVp and GTVn. Radiomics features were respectively extracted for GTV, GTVp, and GTVn using Pyradiomics. Competing risk analysis was used to identify significant features for BM, in which death without BM was considered a competing event. Bootstrapping samples with 500 iterations was performed to train models. Spearman correlations were performed for features that were significant in univariate analysis. Significant correlated features were removed and the ones with the highest hazard ratio (HR) were included in the multivariate model. Area under the receiver operating characteristic curves (AUC-ROC) was performed to assess model predictive performance.

In total, 296 out of 497 patients were eligible, 157 (53%) were male, 112 (37.8%) were squamous cell,160 (54.1%) were stage IIIA, and 73 (24.7%) received immunotherapy. The median age was 66y (IQR 60-72). Within a median follow up of 55.3 months (95% CI 48.0-62.7 months), 180 (60.8%) patients died, 46 (15.5%) developed BM at a median time of 10.8 months. The median overall survival was 29.9 months (95%CI 23.1-36.7). GTV was available for all patients, of which 266 GTVn and 274 GTVp were distinguishable.  In all, 861 features were extracted. Univariate analysis showed that 3 GTV, 4 GTVn, and 2 GTVp features were significantly associated with BM development (p<0.05). Multivariate models showed that Wavelet-HLL-first-order-Median (HR=1.62, 95% CI 1.04-2.51, p=0.03) was a significant GTV feature, Wavelet- HLL-first-order-Median (HR=2.21, 95% CI 1.32-3.68, p=0.002) and Wavelet-HHL-firstorder-Median (HR=1.61, 95% CI 1.05-2.46, p=0.03) were significant GTVn features, Wavelet-LLH-glcm-Idn (p=0.06) and Wavelet-LHH-glcm-Imc1 (p=0.06) were marginally significant GTVn features, Wavelet-HLH-glszm-Gray-Level-Non-Uniformity (HR=1.48, 95% CI: 1.16-1.90, p=0.002) and Wavelet-LLL-glcm-Imc2 (HR=0.69, 95% CI: 0.51-0.94,p=0.02) were significant GTVp features (Table). The AUC of the GTV, GTVn, GTVp model ranged from 0.59-0.68, 0.65-0.78, 0.49-0.64, respectively (Figure).

Radiomics features based on TRT planning CT have predictive value for BM development in stage III NSCLC. GTVn provided more predictive value than GTVp and GTV.