Fig. 2. Extraction of CT radiomics features and association with response.
a, Radiomics feature extraction pipeline using expert segmented thoracic CT scans. Superpixel-based perturbations on original segmentations used for feature selection. b, Three expert CT segmentation examples including lung parenchymal (PC) (top), pleural (PL) (middle) and lymph node (LN) lesions (bottom) representative of n = 187 patients, with the original image, segmentation and randomized contour example. c, Principal-component decomposition distribution of radiomics features for superpixel-based perturbations across tissue sites. The interior box-and-whisker bars show the mean as a white dot, the IQR (25–75%) as a black bar and the minimum and maximum as whiskers up to 1.5 × IQR. P values were obtained from the two-sided Mann–Whitney–Wilcoxon test for n = 2,040 PC perturbations with n = 540 in the responding group and n = 1,500 in the nonresponding group; n = 330 PL perturbations with n = 130 in the responding group and n = 200 in the nonresponding group; and n = 960 LN perturbations with n = 360 in the responding group and n = 600 in the nonresponding group. d, Response prediction performance using LR classifiers for each type of lesion as well as averaging-based patient-level prediction aggregation by averaging (LR Rad-Average) outcomes across all lesions and the multiple-instance learning model. Results with AUC < 0.5 are not shown. The bar height and error bar represent the AUC and associated 95% CI based on DeLong’s method51 for n = 187 and n = 46 patients in the multimodal and validation cohorts, respectively.