Use of quantitative histomorphometrics to classify disease progression in HPV-positive squamous cell carcinoma.

Abstract

73 Background: Oropharyngeal squamous cell carcinoma (SCC) is now recognized as distinct because of its strong association with high-risk human papillomavirus (HPV) and favorable prognosis. Although these tumors have some specific morphologic features, none have been shown to predict prognosis, particularly on small specimens such as biopsies. Methods: We present a novel cluster cell graph (CCG) that is computationally efficient and provides an effective tool to quantitatively describe histopathologic images according to the spatial distribution and clustering of cells. CCG is generated by firstly automatically identifying all cells/nuclei within the histologic image and then identifying nuclear clusters. These clusters are then represented via a single node corresponding to center of mass of the cluster. Any pair of nodes is connected via a link if the probability of such an association is greater than a pre-defined threshold, the probability being calculated as a decaying function of the Euclidean distance between node pairs. Once links have been identified between cluster nodes, subgraphs are constructed as series of topological features defined on each node of the subgraph, i.e., local graph metrics (e.g. clustering coefficients and skewness of edge lengths). Results: We leverage CCG to capture structural characteristics of HPV-related (p16+) oropharyngeal SCC (oSCC) from formalin-fixed, paraffin-embedded tissue microarray images. A series of quantitative histomorphometric descriptors are extracted for each image and used to train a random forest (RF) classifier to distinguish patients with p16+ oSCC who develop disease progression from those that do not. Over 25 runs of 3-fold cross validation in a cohort of 140 studies (116 non-progressors and 26 progressors) the RF classifier yielded a mean accuracy of over 87% in distinguishing between progressors and non-progressors. Conclusions: The ability to quantitatively describe the spatial distribution/clustering of nuclei via CCG allows for building of classifiers that can distinguish between progressors and non-progressors in p16+ oSCC tissue images. The approach should be extensible to outcome prediction for other cancers as well.