Characterization of Lung and Pancreatic Tumors in the Deep Learning Era

Abstract

Using computer-aided diagnosis (CAD) technologies, risk stratification (characterization) of cancers using radiological images may be done more quickly and accurately. As part of precision medicine, tumour characterisation using such methods can also support non-invasive cancer staging, prognosis, and individualised therapy planning. We suggest supervised and unsupervised machine learning techniques in this research to enhance tumour characterisation. We demonstrate considerable improvements using deep learning algorithms for our first method, which is based on supervised learning. We use a 3D convolutional neural network and transfer learning in particular. We next demonstrate how to use a graph-regularized sparse multi-task learning framework to include task-dependent feature representations into a CAD system, which is motivated by the radiologists' interpretations of the scans.