Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Classification of Anatomic Structures by CT-Based Radiomics for Head-Neck Radiotherapy

Y Watanabe1*, N Gopishankar2, A Biswas2, K Rangarajan2, G Rath2, (1) University of Minnesota, Minneapolis, MN, (2) All India Institute Of Medical Sciences


PO-GePV-M-44 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: To show that CT images' radiomics features can classify anatomic structures in the head and neck.

Methods: We obtained 26 sets of CT image data for patients treated at a single institution for head and neck cancers using the IMRT technique. We calculated 174 radiomics feature values by the SIBEX software, which complies with the Image Biomarker Standardization Initiatives (IBSI) protocol. The radiomics calculations were done using three processing parameters: FBN, FBS with 25 and 10 bin widths. For this study, we chose several representative anatomical structures, i.e., 27 gross tumor volumes (GTV), a total of 47 right and left parotids, 25 mandibles, one brainstem, two eyes, and two sub-mandibles, which were segmented by attending radiation oncologists for treatment planning. First, we applied the principal component analysis (PCA) method to the feature values. Using the first two principal components, we studied if we can classify the anatomical structures by the k-means clustering method. Furthermore, the quantitative differences of 174 radiomics features were visualized by plotting heatmaps and correlation plots of all feature values among the structures.

Results: The heatmaps of 174 normalized feature values of mandible showed a clear difference from GTV and parotids. The difference between GTV and parotids were also observed. The correlation plots showed the differences in feature values among GTV, parotid, and mandible. The first two principal components' two-dimensional plot indicated the five anatomical structures in the figure's specific locations. The radiomics features calculated with FBS=10 showed the clearest classification of GTV, parotid, and mandible among three different feature calculation methods. The clustering method could successfully classify those three structures.

Conclusion: We could classify head-neck tumors, parotids, and mandibles by radiomics features. The results imply that the radiomics can decipher the differences of biological functions, microenvironment, and cellular structures from CT images.



    Image Analysis, Quantitative Imaging, Classifier Design


    IM- Dataset Analysis/Biomathematics: Machine learning

    Contact Email