MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Unsupervised Machine Learning of Metabolic Response From Radiomic Expression of Oropharyngeal Cancers

K Lafata*, Y Chang, C Wang, Y Mowery, I Vergalasova, J Liu, D Brizel, F Yin, Duke University, Durham, NC

Presentations

(Sunday, 7/12/2020) 1:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose: To identify computational imaging biomarkers of early metabolic response in patients undergoing definitive radiotherapy for oropharyngeal cancer (OPC).

Methods: Seventy-two patients were enrolled in a prospective clinical trial to receive definitive radiotherapy (70Gy) for OPC. PET/CT images were acquired both prior to treatment and two weeks into treatment (i.e., after 20Gy). All patients were scanned on the same PET/CT imaging system. The gross tumor volume at the primary tumor site was manually segmented on CT and transferred to PET, from which 55 quantitative radiomic features were extracted as potential biomarkers for early therapeutic response. An unsupervised machine learning algorithm was used to cluster patients into groups based solely on the intrinsic properties of their radiomic signatures. Clinical outcomes were prospectively collected following therapy to identify evidence of disease progression. The primary clinical endpoints were cancer recurrence and progression free survival (PFS). The clustered radiomics data were naïvely compared to cancer recurrence, from which Kaplan-Meier estimators were derived in a cluster-by-cluster fashion.

Results: While no relationship was observed between pre-treatment radiomics and clinical outcomes, unsupervised clustering of intra-treatment radiomics revealed 3 patient clusters associated with cancer recurrence (p=0.0256, ?² test) and PFS (p=0.0053, log-rank test). In general, clusters associated with poor prognosis were characterized by metabolic heterogeneity and high FDG uptake. In contrast, clusters associated with favorable prognosis were characterized by metabolic homogeneity and low FDG uptake. Clusters representing the largest separation of radiomic signal resulted in a hazard ratio of 7.53 (CI=2.54-22.3; p=0.0201, log-rank test).

Conclusions: By combining techniques from computer vision and unsupervised machine learning, we have developed a parsimonious approach to quantifying early metabolic response based on the intrinsic properties of radiomics data. Our results collectively suggest that intra-treatment PET radiomic expression is a prognostic biomarker for OPC, regardless of the baseline metabolic characteristics of the disease.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email