Radiomic feature stability across 4D respiratory phases and its impact on lung tumor prognosis prediction

Radiomic analysis has recently demonstrated versatile uses in improving diagnostic and prognostic prediction accuracy for lung cancer. However, since lung tumors are subject to substantial motion due to respiration, the stability of radiomic features over the respiratory cycle of the patient needs to be investigated to better evaluate the robustness of the inter-patient feature variability for clinical applications, and its impact in such applications needs to be assessed. A full panel of 841 radiomic features, including tumor intensity, shape, texture, and wavelet features, were extracted from individual phases of a four-dimensional (4D) computed tomography on 20 early-stage non-small-cell lung cancer (NSCLC) patients. The stability of each radiomic feature was assessed across different phase images of the same patient using the coefficient of variation (COV). The relationship between individual COVs and tumor motion magnitude was inspected. Population COVs, the mean COVs of all 20 patients, were used to evaluate feature motion stability and categorize the radiomic features into 4 different groups. The two extremes, the Very Small group (COV≤5%) and the Large group (COV>20%), each accounted for about a quarter of the features. Shape features were the most stable, with COV≤10% for all features. A clinical study was subsequently conducted using 140 early-stage NSCLC patients. Radiomic features were employed to predict the overall survival with a 500-round bootstrapping. Identical multiple regression model development process was applied, and the model performance was compared between models with and without a feature pre-selection step based on 4D COV to pre-exclude unstable features. Among the systematically tested cutoff values, feature pre-selection with 4D COV≤5% achieved the optimal model performance. The resulting 3-feature radiomic model significantly outperformed its counterpart with no 4D COV pre-selection, with P = 2.16x10-27 in the one-tailed t-test comparing the prediction performances of the two models.