Abstract
Risk stratification (characterization) of tumors from radiology images can bemore accurate and faster with computer-aided diagnosis (CAD) tools. Tumorcharacterization through such tools can also enable non-invasive cancerstaging, prognosis, and foster personalized treatment planning as a part ofprecision medicine. In this study, we propose both supervised and unsupervisedmachine learning strategies to improve tumor characterization. Our firstapproach is based on supervised learning for which we demonstrate significantgains with deep learning algorithms, particularly by utilizing a 3DConvolutional Neural Network and Transfer Learning. Motivated by theradiologists' interpretations of the scans, we then show how to incorporatetask dependent feature representations into a CAD system via agraph-regularized sparse Multi-Task Learning (MTL) framework. In the secondapproach, we explore an unsupervised learning algorithm to address the limitedavailability of labeled training data, a common problem in medical imagingapplications. Inspired by learning from label proportion (LLP) approaches incomputer vision, we propose to use proportion-SVM for characterizing tumors. Wealso seek the answer to the fundamental question about the goodness of "deepfeatures" for unsupervised tumor classification. We evaluate our proposedsupervised and unsupervised learning algorithms on two different tumordiagnosis challenges: lung and pancreas with 1018 CT and 171 MRI scans,respectively, and obtain the state-of-the-art sensitivity and specificityresults in both problems.