Bayesian classifiers of solid lesions with dynamic CT: Integrating enhancement density with washout density and delay interval

Solid lesions emerge within diverse host tissue environments, making their diagnosis a challenge on the basis of non-invasive techniques alone. Various techniques have been developed to transform images into quantifiable feature sets producing summary statistics that describe enhancement distributions of solid masses. Relying on empirical distributional summaries, conventional approaches to cancer radiomics analysis are limited by the application of regression or machine learning algorithms to correlated feature sets. Moreover, often relying on summary statistics that derive from a single scan, current analytical approaches ignore temporal patterns that describe washout across multiple scans with contrast. Motivated by the diagnosis of adrenal masses on the basis of dynamic contrast enhanced computed tomography, this article presents novel statistical methodology for formulating similarity networks that integrate the entire enhancement and washout distributions of a delineated region of interest (ROI) with the duration of the delay scan interval. Applying consensus clustering to the network, we demonstrate unsupervised learning, revealing five discrete patterns of the combination of non-contrast enhancement distribution and contrast washout. Additionally, we demonstrate how the resultant network may be used to train a supervised Bayesian classifier based on the concept of partial exchangeability. When applied to predict the true pathology-verified malignancy status of adrenal lesions in our study, classification using Bayesian predictive probabilities deriving from the similarity network yielded an area under the ROC curve of 0.908 outperforming prediction with conventional regression analysis based on summary statistics of the enhancement densities, which yielded only 0.828.