Using a Noisy U-Net for Detecting Lung Nodule Candidates

Chest computed tomography is an important method for detecting lung cancer. Because early lung cancer nodules are small, making them difficult to detect, the current automatic lung cancer detection system can easily lead to missed diagnosis when detecting these nodules; therefore, accurate detection of early lung cancer nodules is crucial for improving the lung cancer cure rate. To reduce the missed diagnosis rate of early nodules in the detection system, it is necessary to optimize the extraction steps of candidate nodules. Based on the improvement of U-Net, this paper proposes noisy U-Net (NU-Net), which can enhance the neural network's sensitivity to small nodules by adding a special noise to the hidden layers in training. The neural network is trained using the LUng Nodule Analysis 2016 dataset and the Alibaba Tianchi Lung Cancer Detection Competition dataset. The comparative experiment between the U-Net and NU-Net reveals that the proposed algorithm's sensitivity to small nodules with diameters of 3-5 mm (97.1%) is greater than the U-Net value (90.5%).