Spatial Continuity and Volumetric Convolutions for Lung Nodule Precision

Abstract

Computer-aided detection (CADe) systems have been developed and automated to reduce cognitive fatigue in diagnostic radiology and help verify the presence of pulmonary nodules. The architectural design paradigm, however, in the form of a trade-off between deep multi-planar 2D slice feature recycling and raw native 3D volumetric convolutions, is still a very hot topic. This paper proposes a strong baseline architecture which directly compares a multi-planar deep feature-reuse network with DenseNet169 and a custom native volumetric convolutional network based on AlexNet3D. We use an automated 360° rotation verification engine with a center exactly at the center of mass of the nodule to validate its coordinates, evaluated on the standardized LUNA16 dataset. In terms of quantitative results, DenseNet169 outperforms AlexNet3D with an Area Under the Curve (AUC) of , sensitivity of  and specificity of  while AlexNet3D had an AUC of , sensitivity of  and specificity of . This condition is the ability of computer-aided detection (CAD) tools to achieve high accuracy in benchmark test. This condition is the ability of computer-aided detection (CAD) tools to demonstrate high accuracy when tested against a standard set.