Detection of Pulmonary Nodules in CT Images Based on 3D Multi-Scale Retina Vnet

Currently, one of the most common and lethal malignancy is lung cancer, the early diagnosis and treatment of which are critical for prognoses and improving the survival rates of patients. Pulmonary nodules are the symptoms of lung cancer. And clinical distinguish between benign and malignant nodules is measuring their volumes and forms after segmentation. However, it is very inefficient to perform lung nodule segmentation by conventional manual methods. In this paper, a lung nodule identification system based on Retina VNet is suggested. The network inherits the structure of the original VNet while introducing Retina Net to extract multi-scale features of pulmonary nodule images. A deep supervision technique is also used for feature tuning, which may significantly increase the model's detection performance. The LIDC-IDRI lung nodule public dataset is used to assess the model's performance in this research. Results showed that the method proposed in this article is superior to the existing detection methods in lung nodule detection performance (especially in detection of lung nodules of various sizes). Average Precision can reach up to 88.7%.