YOLO_v3-Based Pulmonary Nodules Recognition System

Abstract

The incidence of pneumonia and lung cancer is high in China. Lung CT images are widely used in the screening and adjuvant treatment of lung diseases due to their advantages of a thin layer, high definition, and low noise. The manual film-reading method has strong subjectivity, depends on the doctor’s experience, and now hospitals produce a large number of lung CT images every day. The manual film-reading method is inefficient. The machine can help doctors to do lesion screening, auxiliary diagnosis, and treatment. Conventional machine learning methods have been applied to the recognition of pulmonary nodules in CT images. Still, these methods need to manually extract features that are not comprehensive or proper, resulting in misdiagnosis and missed diagnosis. These methods enable to rapidly detect lung nodules, which is terrible for patients to receive treatment in a timely manner. With the development of deep learning technology, Convolutional Neural Networks (CNNs) have been widely used in recognizing images, such as facial identification, character recognition, car plate recognition, etc. Its ability to solve computer vision problems has won the approval in some natural scene tasks. In this paper, the YOLO_v3 is used for feature extraction and classification, and residual network (ResNet), one of the classical CNN models, is used to decrease the false-positive rate, which improves the detection accuracy of pulmonary nodules in CT images.