Co-occurring gland tensors in localized cluster graphs: Quantitative histomorphometry for predicting biochemical recurrence for intermediate grade prostate cancer

Abstract

Quantitative histomorphometry (QH), computational tools to analyze digitized tissue histology, has become increasingly important for aiding pathologists in assessing cancer severity. In this study, we introduce a novel set of QH features utilizing co-occurring gland tensors (CGT) in localized cluster graphs to quantitatively evaluate prostate cancer (CaP) histology. CGTs offer three main advantages over previous QH features: 1) gland tensors represent a novel measurement that has been anecdotally described as one of interest, but never quantitatively modeled, 2) CGTs extract measurements based on local rather than global glandular networks, constructed using cluster graphs, and 3) second order statistical features (energy, homogeneity, energy, and correlation) obtained from a co-occurrence matrix capture the spatial interactions of gland tensors in the image. We extract 4 CGT features from 56 regions across 40 intermediate grade CaP patients and evaluated the ability of CGT features to predict biochemical recurrence (BCR) within 5 years of radical prostatectomy. Intermediate Gleason score 7 cancers represent the predictive borderline for BCR cases, where 50% of cases develop BCR. We found that CGT features outperformed 5 different sets of QH features, previously shown to be effective in CaP grading, when evaluated via a Random Forest classifier (66% accuracy for CGT features versus 55% for the next closest QH feature set), all comparisons being statistically significant.