Controlling False-Positives in Automatic Lung Nodule Detection by Adding 3D Cuboid Attention to a Convolutional Neural Network

Abstract

Background and ObjectiveFalse-positive diagnoses often occur during pulmonary nodule detection, which directly affects the prognosis of patients. The development of intelligent computing and the wide application of deep learning technology have brought about the intelligent detection of lung nodules, which has greatly improved the detection efficiency and detection accuracy. MethodsThis paper explores a convolutional neural network filled with a 3D cuboid attention module for lung nodule detection comprehensively encompassing nodules in the cross, coronal, and sagittal planes of CT sections. First, 3D cuboid attention module extracts image features in order from the plane, space, and volume directions of CT slices (PSV-AM). Second, the model adopts a structure named UPSV-Net of 9 modules to expose the lesion area as much as possible. The PSV-AM is embedded into the network to guide it in extracting the structural features of nodules. At the same time, the features extracted from the shallow and deep layers are associatively concatenated. Thirdly, on the basis of the convolution module shared from the detection branch, deepening the convolution path to form the true and false nodule classification network, which is committed to reducing the false positive rate. ResultsThe model study was validated on the 2016 Lung Nodule Analysis challenge dataset. On the test bench, the detection accuracy is 96.15%, and the sensitivity is 97.6%. Subsequently, the sensitivity to the true and false nodules reached 97.61% and the average false-positive detection rate (FPDR) was 0.893, which was comparable to the diagnosis results of clinicians.