Supervised and Unsupervised Tumor Characterization in the Deep Learning Era.

Computer Aided Diagnosis (CAD) tools are often needed for fast and accurate detection, characterization, and risk assessment of different tumors from radiology images. Any improvement in robust and accurate image-based tumor characterization can assist in determining non-invasive cancer stage, prognosis, and personalized treatment planning as a part of precision medicine. In this study, we propose both supervised and unsupervised machine learning strategies to improve tumor characterization. Our first approach is based on supervised learning for which we demonstrate significant gains in learning algorithms, particularly Convolutional Neural Network (CNN), by utilizing completely 3D approach and transfer learning to address the requirements of volumetric and large amount of training data, respectively. Motivated by the radiologists' interpretations of the scans, we then show how to incorporate task dependent feature representations into a CAD system via a graph regularized sparse Multi-Task Learning (MTL) framework. In the second approach, we explore an unsupervised scheme in order to address the limited availability of labeled training data, a common problem in medical imaging applications. Inspired by learning from label proportion (LLP) approaches, we propose a new algorithm, proportion-SVM, to characterize tumor types. In this second approach, we also seek the answer to the fundamental question about the goodness of deep features for unsupervised tumor classification. Finally, we study the effect of unsupervised representation learning using Generative Adversarial Networks (GAN) on classification performance. We evaluate our proposed approaches (both supervised and unsupervised) on two different tumor diagnosis challenges: lung and pancreas with 1018 CT and 171 MRI scans respectively.