(Sunday, 7/29/2018) 4:00 PM - 4:55 PM
Room: Room 205
Purpose: We investigated whether radiomics features on contrast enhanced computed tomography (CT) are correlated with biological characteristics in head-and-neck squamous cell cancer (HNSCC) patients.
Methods: We retrospectively analyzed 81 HNSCC patients with high quality pre-treatment CT scans from The Cancer Imaging Atlas (TCIA). Primary tumors were manually contoured by a single physician and reviewed by two experienced radiologists. Using the CERR radiomics toolbox, 676 imaging features were extracted from segmented regions of interest. Unsupervised hierarchical consensus clustering was used to uncover potential subgroups of HNSCC using 80% sample and 80% feature subsampling. Based on Bonferroni correction, significant driving imaging features that separate the identified subgroups were found. Hierarchical clustering was used to remove redundant features among the significant imaging features. The correlation between the subgroups/imaging features and tumor infiltrating lymphocytes (TILs) from The Cancer Genome Atlas (TCGA) was investigated.
Results: The subsite of the cohort consisted of oral cavity (n=40), larynx (n=28), and oropharynx (n=13). Consensus clustering led to two subgroups with 41 and 40 samples in subgroups A and B, respectively. All oropharyngeal tumors were clustered together in subgroup A with 14 oral cavity and 14 larynx tumors (p=0.0002). Regarding HPV status, 14 tumors were HPV+; subgroups A and B contained 12 and 2 HPV+ tumors, respectively, with 10 in oropharyngeal tumors. In total, 20 significant imaging features were found after removing redundant features. For activated CD4+ and CD8+ T cells, average z-scores of gene expressions from the corresponding metagenes were compared with the 20 imaging features, resulting in 15 and 0 significant features (p<0.05) in CD4+ and CD8+ T cells, respectively. The correlation between subgroups and TILs showed borderline significance with p=0.12 in CD4+ T cell.
Conclusion: We showed that radiomics features on CT scans can uncover potential subgroups of HNSCC and these subgroups correlate with immune activation.
Keywords
Texture Analysis, Modeling, Image Analysis
Taxonomy
IM- Dataset analysis/biomathematics: Machine learning
Contact Email
