(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: This work presents radiomics network graphs to understand the interconnections between radiomics features.
Methods: The methodology consists of deriving the “N” number of feature sets from a dataset. The feature sets can be bootstraps of the dataset in case of a sufficiently large sample size. Otherwise, the feature sets can be derived by perturbing images and segmentations from the original dataset. Network graphs are then computed by averaging partial correlation coefficients across feature sets. The average and standard deviation of partial correlation coefficients indicate sparsity and robustness respectively. A graphical user interface was created to visualize the network graph. The interface allows controlling the sparsity of the network by thresholding partial correlations. The thickness of the network edges represents the partial correlation. Nodes can be highlighted based on (i) feature class, (ii) robustness and (iii) signatures. The graphical interface was implemented as a part of the Radiomics toolbox in the Computational Environment for Radiological Research (CERR) software. CERR allows generation of radiomics signatures as a part of its library of model implementations which are then visualized on network graphs.
Results: 420 patients from the TCIA dataset for lung cancer were used to create radiomics network graph. 50 feature sets were derived by perturbing original images and segmentations. A dense network was initially generated to get an idea of interconnections between features. The network was further split into its components to understand feature clusters. Features from Aerts et al signature were visualized on the network.
Conclusion: Software to generate radiomics network graphs is distributed with radiomics toolbox within CERR at https://www.github.com/cerr/CERR. Radiomics network graphs help understand the robustness of features and the published signatures. They can be used to identify robust and complimentary features for further modeling.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
