How transferable are the deep features from false positive reduction network for lung nodule detection in CT to malignancy prediction?

Gwangbeen Park, Seungho Lee, Hyunho Park, Kyu-Hwan Jung

Abstract: Achieving highly generalizable deep features from data is a fundamentally important problem in various tasks. This work presents experimental results that we can use features from a false positive reduction network for lung nodule detection as generalizable nodule features. Feature visualization with t-SNE and nodule feature based similar nodule search results show that these features have discriminative malignancy and shape information even though the network was only trained to classify nodules from non-nodules.

Author Affiliation: VUNO Inc.

Keywords: lung nodule, CBIR