Risk stratification of lung nodules using 3D CNN-based multi-task learning

Summary

Hussein, Cao, Song and Bagci describe a method for estimating lung-nodule malignancy scores from computed tomography (CT) scans. The approach uses a three-dimensional convolutional neural network (3D CNN) to represent volumetric scan information, transfer learning to support feature learning, and multi-task learning to combine features associated with six nodule attributes. A graph-regularized sparse model is used to account for potential disagreement among radiologists when scoring attributes. The authors report evaluating the approach on a publicly available dataset comprising 1,018 scans. This is a research evaluation; the paper does not establish prospective clinical validation or readiness for clinical use.