Deep Learning on Point Clouds for False Positive Reduction at Nodule Detection in Chest CT Scans

This paper focuses on a novel approach for false-positive reduction (FPR) of nodule candidates in Computer-Aided Detection (CADe) systems following the suspicious lesions detection stage. Contrary to typical decisions medical image analysis, proposed considers input data not as 2D or 3D image, but rather point cloud, and uses deep learning models clouds. We discovered that cloud require less memory are faster both training inference compared traditional CNN 3D, they achieve better performance do impose restrictions size i.e. no candidate. propose an algorithm transforming CT scan cloud. In some cases, volume candidate can be much smaller than surrounding context, example, case subpleural localization nodule. Therefore, we developed sampling points from constructed region. The is able guarantee capture context information part designed set up experiment creating dataset open LIDC-IDRI database feature FPR task, herein described detail. Data augmentation was applied avoid overfitting upsampling method. Experiments were conducted with PointNet, PointNet++, DGCNN. show outperforms baseline resulted 85.98 FROC versus 77.26 models. compare our published SOTA demonstrate even without significant modifications it works at appropriate level LUNA2016 shows LIDC-IDRI.