Deep Radiomics Feature Analysis for HPV-Based Stratification of Patients with Oropharyngeal Cancer Using Kohonen Self-Organizing-Map Technique
Bagher-Ebadian H, Lu M, Siddiqui F, Ghanem AI, Wen N, Wu Q, Liu C, Movsas B, and Chetty IJ. Deep Radiomics Feature Analysis for HPV-Based Stratification of Patients with Oropharyngeal Cancer Using Kohonen Self-Organizing-Map Technique. J Med Phys 2019; 46(6):e400.
J Med Phys
Purpose: To perform deep and unsupervised characteristic analysis of radiomics features extracted from primary tumors of pre-treatment contrastenhanced computed-tomography (CE-CT) images of patients with oropharyngeal cancers to construct an unsupervised classifier to stratify patients on the basis of Human-Papilloma-Virus (HPV) status. Methods: One-hundredeighty- seven patients with oropharyngeal cancers with known HPV-status (confirmed by immunohistochemistry-P16-protein testing) were retrospectively studied. Radiomics features (n = 172) were extracted from CE-CT images of gross-tumor-volume. Levene and Kolmogorov-Smirnov's tests with absolute-biserial-correlation were used to identify discriminant features between HPV+ and HPV- groups. Given the discriminant features, an optimal Kohonen-Self-Organizing-Map (K-SOM; optimal-topology: [9 × 9]- neurons) was constructed using an unsupervised-competitive-learning algorithm. The dataset was split (100-iterations) into training and test (67% vs 33% respectively) cohorts using a bootstrapping-permutation-sampling technique. In each iteration, the K-SOM was trained and evaluated by the training and test cohorts respectively. K-SOM neighbor-weight-distance and their corresponding hit/(winner-frequency) maps were computed for 100-iterations. Regions pertaining to the dominant characteristics of the discriminant features were identified in the K-SOM space and used to calculate stratification power of the classifier. A correlation matrix was calculated for the dominant regions of the K-SOM weight maps (12 maps) and used for further feature characteristic interpretation. Results: Three feature categories were found to be discriminant between the two groups: Discrete-Orthonormal-Stockwell- Transform (DOST), Gray-Level-Co-Occurrence-Matrix (GLCM), and Morphology. The K-SOM revealed four dominant regions associated with the HPV statuses: high-certainty-HPV-, low-certainty-HPV-, low-certainty- HPV+, and high-certainty-HPV + . Average Area-Under-Receiver-Operating- Characteristic of the optimal K-SOM was 0.88 ± 0.06. Conclusion: This study presents an unsupervised-intuitive deep radiomics feature analysis that provides a data-driven feature characterization. It not only offers high performance HPV-based patient stratification, but also reveals a full-spectrum of feature gains, variations, similarities, and dissimilarities which are key parameters in building different classifiers for HPV-based stratification of patients with oropharyngeal cancer.