Click here to

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

Radiomic Analysis of Limited-Sized Data Using Unsupervised Methods: Prediction of HPV Status in Head and Neck Cancer

J Oh1*, A Apte1, A Iyer1, A Tannenbaum2, J Deasy1, (1) Memorial Sloan Kettering Cancer Center, New York, NY, (2) Stony Brook University


PO-GePV-M-11 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: Working with limited-sized data is a common challenge in medicine. We investigated a promising unsupervised approach to radiomics that combines unsupervised optimal mass transport (OMT) on a network of radiomic features. tSNE (t-stochastic neighborhood embedding) was then used to cluster and visualize the results in 2-D with the novel Wasserstein metric. We demonstrate this approach on the important problem of predicting HPV status in head-and-neck squamous cell cancer (HNSCC) patients with contrast enhanced computed tomography (CT) images.

Methods: Radiomic analysis was performed on 77 HNSCC patients with pre-treatment CT scans downloaded from The Cancer Imaging Archive (TCIA). Using the CERR radiomics toolbox, 67 radiomic features were extracted from primary tumors manually contoured on all evaluable artifact-free slices. First, a partial correlation network was constructed, where a link between two nodes was determined using a partial correlation coefficient after conditioning on all other features. To further remove potential false positive links, a lasso-type regularization was applied to the resultant network. On the largest connected network component, an OMT metric called the W₁ Wasserstein distance was computed for all pairs of samples, and then the resultant distance matrix was input into tSNE to map the samples to a low-dimensional space.

Results: The cohort had 13 HPV positive tumors. The resultant Wasserstein distance matrix computed in the largest network component with 10 radiomic features was fed into tSNE, using the OMT-based distance as a cost function. In the mapped two-dimensional space, tumors with HPV positivity were better clustered with a Chi-squared p-value of 0.0008 than the tSNE result with the Euclidean distance.

Conclusion: We demonstrated that radiomic features represented as a network can predict HPV positivity in head and neck cancer with good accuracy, employing tSNE coupled with an OMT based metric. The unsupervised result could further be refined using supervised methods.



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email