Session Item

Radiomics is increasingly used to implement clinically-based prognostic models for non-small cell lung cancer (NSLCL). However, no evidence supports the choice of specific imaging pre-processing methodologies. Admittedly, dedicated investigations could contribute to refining both the reproducibility and the performance of radiomic studies. As the volume of the primary tumor is a well-recognized prognosticator, our purpose is to assess how pre-processing may impact the feature-volume dependency in computed tomography (CT) images of NSCLC patients treated with radiotherapy.

Images were retrieved from the publicly available repository NSCLC-Radiomics of The Cancer Imaging Archive (TCIA). Four hundred eighteen images were included in the analysis following manual inspection and editing of the segmentations; nodal disease- if any- was not included. Pyradiomics was used to extract 93 features; which were grouped as follows: first-order, shape-based (3D), shape-based (2D), gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), neighboring gray tone difference matrix (NGTDM) and gray level dependence matrix (GLDM). Twenty built-in pre-processing methods (filters) were applied, including wavelet and its possible permutations, Laplacian of gaussian, and local binary pattern (LPB); each feature except those belonging to the shape category was computed once per filter, and on the original CT image. The Spearman correlation coefficient (ρ) was used; with thresholds of ≥0.7 and ≤0.5 defining strong and weak correlations, respectively.

Overall, features of the GLCM category were the least correlated with volume (ρ= 0.5). The feature/volume correlation was found to be filter-dependent: the highest correlation was found when lpb-3D-m1 was applied (ρ= 0.82), while the lowest correlations with volume were identified for the HHL and HHH-wavelet filters, and for the exponential method (ρ= 0.35, 0.30 and 0.18, respectively). These results were confirmed when features computed per each pre-processing modality were compared to the original image. An overview of the results is displayed in Figure 1.

Our results support the hypothesis that pre-processing does impact features values; and provide a proof of concept that further standardization is warranted for radiomic studies. Further- currently ongoing analyses- will focus on how these findings impact the performance of radiomic-based survival models.