Click here to


Are you sure ?

Yes, do it No, cancel

Radiomics Feature Robustness Under Different Image Perturbation Combinations and Intensities: A Study On Nasopharyngeal Carcinoma CT Images

J Zhang1, X Teng1*, Z Ma1, T Yu1, S Lam1, F Lee2, K Au2, W Yip2, J Cai1, (1) The Hong Kong Polytechnic University, Hung Hom, Kowloon, HKSAR, (2) Queen Elizabeth Hospital, HKSAR


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: To provide guidance on determining image perturbation intensity and combinations for optimum model generalizability in radiomics study.
Methods: Image perturbations are simulated by combinations of rotation(R), noise addition(N), translation(T), volume adaption(V) and supervoxel-based contour randomization(C). We improved the previously proposed perturbation method by quantifying and fine tuning the elevated noise level for N and randomized contour similarity to the original for C. 21 image perturbation groups (40 combinations in each) were applied to CT images from 76 nasopharyngeal carcinoma patients. 1302 radiomics features on GTV were calculated for each perturbation combination. The feature robustness susceptibility from each perturbation mode were compared among radiomics feature categories. Both individual and collective change of feature robustness under varying perturbation combinations and intensities were statistically analyzed.
Results: Wavelet features have the largest portion of unrobust features under all perturbation groups compared with original and log-sigma features. For example, 34.9% wavelet features, 2.2% log-sigma features, and 4.7% original features were classified as poorly robust under the 5-mode perturbations. Less than 1% of original and log-sigma features are non-excellent under R and T, indicating their minimum dependency on resample variations. N causes a large portion of original (17.0%) and wavelet (21.1%) features to be poorly robust but have minimum effect on log-sigma features (0.4%) due to the smoothing filter. Both increasing perturbation intensity and including more perturbation modes into combinations deteriorate feature stability.
Conclusion: Our results on feature robustness behavior under varying perturbation intensities and combinations provide guidance on image perturbation simulation design for balanced model generalizability and predictive power. They also serve as a reference feature robustness record when evaluating the existing model performance. The optimum perturbation parameters that best simulate the clinical scenario and generalization on other image modalities and treatment sites will be explored in the next phase of our study.


Perturbation Factor, Modeling, Feature Selection


IM- CT: Radiomics

Contact Email