306 Correlation between radiomic features extracted from CT images of non small cells lung cancer (NSCLC) and lymph node status: Preliminary results

Purpose To extract radiomic features from CT images of non small cells lung cancer (NSCLC) patients, assess their repeatability and reproducibility, and investigate their possible correlation with lymph node status (pN). Methods CT images of 481 patients affected by NSCLC were considered, selecting only images acquired at our Institution with 120 kV and 2.5 mm slice thickness. The 3D Regions Of Interest (ROIs) delineating the nodule were drawn by a Radiologist and uploaded in the IBEX tool with the CT images for radiomic features calculation (1419 for each ROI). Features repeatability was investigated with a phantom study. Only features exhibiting Coefficient Of Variation ANOVA was performed to identify the features significantly affected by acquisition parameters as pixel size, mAs, contrast medium type, or by ROI volume. Hierarchical clustering was performed to group the remaining repeatable and reproducible features into clusters, maximizing intra-cluster and minimizing inter-cluster correlation. Wilcoxon rank-sum test was performed to investigate correlation between each feature and pN. For each cluster, the feature yielding highest significant correlation (if present) was considered representative. For comparison, correlation was also investigated between pN and the not-radiomic Signal-to-Noise and Contrast-to-Noise Ratio parameters (SNR, CNR) calculated for each ROI. Results 203 of the 481 patients entered the Radiomic analysis; 520 out of the 1419 extracted features were selected after repeatability and robustness assessment. 15 clusters represented the original dataset without redundancy (Figure); 4 of them contained at least one feature with significant p value at Wilcoxon test (Table), indicating high probability to be correlated with pN. No significant correlation was found between non-radiomic parameters (SNR, CNR) and pN. Conclusion Despite preliminary, the present results evidence that – differently from standard quantitative parameters like SNR and CNR – radiomic features describing tumour texture and shape may predict information about lymph node status.