MIXCAPS: A capsule network-based mixture of experts for lung nodule malignancy prediction

Lung cancer is among the most common and deadliest cancers with a low 5-year survival rate. Timely diagnosis of lung is, therefore, paramount importance as it can save countless lives. In this regard, Computed Tomography (CT) scan widely used for early detection cancer, where human judgment currently considered gold standard approach. Recently, there has been surge interest on development automatic solutions via radiomics, human-centered subject to inter-observer variability highly burdensome. Hand-crafted serving radiologist assistant, requires fine annotations pre-defined features. Deep learning radiomics solutions, however, have promise extracting useful features their own in an end-to-end fashion without having access annotated boundaries. Among different deep models, Capsule Networks are proposed overcome shortcomings Convolutional Neural (CNNs) such inability recognize detailed spatial relations. networks so far shown satisfying performance medical imaging problems. Capitalizing success, study, we propose novel capsule network-based mixture experts, referred MIXCAPS. The MIXCAPS architecture takes advantage not only network’s capabilities handle small datasets, but also automatically splitting dataset through convolutional gating network. enables network experts specialize subsets data. Our results show that outperforms single network, CNN, CNNs, ensemble networks, average accuracy 90.7%, sensitivity 89.5%, specificity 93.4% area under curve 0.956. experiments relation between gate outputs couple hand-crafted features, illustrating explainable nature To further evaluate generalization architecture, additional brain tumor performed showing potentials tumors related other organs.