Liver tissue classification in patients with hepatocellular carcinoma by fusing structured and rotationally invariant context representation

Abstract

This work addresses multi-class liver tissue classification from multi-parameter MRI in patients with hepatocellular carcinoma (HCC), and is among the first to do so. We propose a structured prediction framework to simultaneously classify parenchyma, blood vessels, viable tumor tissue, and necrosis, which overcomes limitations related to classifying these tissue classes individually and consecutively. A novel classification framework is introduced, based on the integration of multi-scale shape and appearance features to initiate the classification, which is iteratively refined by augmenting the feature space with both structured and rotationally invariant label context features. We study further the topic of rotationally invariant label context feature representations, and introduce a method for this purpose based on computing the energies of the spherical harmonic decompositions computed at different frequencies and radii. We test our method on full 3D multi-parameter MRI volumes from 47 patients with HCC and achieve promising results.