Effective Representation of Three-Dimension Nodules for False-Positive Reduction in Pulmonary Nodule Detection

The convolutional neural networks (CNNs) can learn features representation from large amounts of training data, and it has achieved remarkable successes in image processing. However, due to expensive expert annotation and privacy issues, the lack of sufficient training data limits the application of CNNs, especially for 3D CNNs, in medical images. In this paper, we decomposed a 3D sample into a set of 2D images based on a series of sequential uniformly-distributed viewpoints and used some "effective" 2D images of them to train a 2D CNN for the false-positive reduction in pulmonary nodule detection with fewer model parameters and less computational burden. The 2D images from different representative viewpoints would provide more complete and independent information than simplifying a 3D nodule into several orthogonal planes. And the non-nodules appear clearly as non-circular linear structures in the "effective" 2D images, which enable us well to distinguish nodules from false positives. Our method was evaluated on 888 CT scans from the database of the LUNA16 challenge. Compared with other methods by using 2D CNNs, our proposed method achieves the highest competition performance metric (CPM) score in the false-positive reduction track. Compared to published work using 3D CNNs which needs significantly larger training data, our method achieves comparable performance by using only about 20% of training data. The "effective" 2D images from representative viewpoints augment the database for training a 2D CNN and provide more crucial information of 3D nodules and would improve the performance of false-positive reduction.