3D Convolutional Neural Networks Fusion Model for Lung Nodule Detection onClinical CT Scans

Abstract

Automatically accurate pulmonary nodule detection plays an important role in lung cancer diagnosis and early treatment. We propose a three-dimensional (3D) Convolutional Neural Networks (ConvNets) fusion model for lung nodule detection on clinical CT scans. Two 3D ConvNets models are trained separately without any pre-training weights: One trained on the LUng Nodule Analysis 2016 dataset (LUNA) and additional augmented data to learn the nodules’ representative features in volumetric space, which may cause overfitting problems meanwhile, so we train another network on original data and fuse the results of the two best-performing models to reduce this risk. Both use reshaped objective function to solve the class imbalance problem and differentiate hard samples from easy samples. More importantly, 335 patients’ CT scans from the hospital are further used to evaluate and help optimize the performance of our approach in the real situation, and we develop a system based on this method. Experimental results show a sensitivity of 95.1% at 8 false positives per scan in Free Receiver Operating Characteristics (FROC) curve analysis, and our system has a pleasing generalization ability in clinical data.