Research And Implementation Of Pulmonary Nodule Detection Method In CT Image Based On YOLOv3 Network

Lung cancer is a malignant tumour and most patients are already in the middle to late stages when it is detected,resulting in a low rate of subsequent cure,and the incidence of lung cancer is still high.The early clinical diagnosis of lung cancer is mainly determined by the shape of the nodes in the lungs,but as lung nodules in the lungs vary in shape and are very unevenly distributed.Relying solely on the physician to make the diagnosis can add to the workload and affect the effectiveness of the diagnosis.The main purpose of this paper is to combine image processing and target detection algorithms to design a lung nodule detection network to automate lung nodule detection.Firstly,CT image segmentation is performed to obtain more easily distinguishable parenchymal lung regions;then an improved target detection algorithm is used to identify and locate nodules for analysis,and finally an automated detection system is designed to assist doctors in medical diagnosis.The specific work is as follows.(1)The conventional lung parenchyma segmentation method cannot extract the lung parenchyma region completely,which will affect the subsequent lung parenchyma detection.Therefore,this paper proposes a lung parenchyma segmentation algorithm based on the OTSU method combined with alternating morphological operations,which can segment the lung parenchyma and the surrounding region,obtain a better similarity coefficient and obtain a more complete lung parenchyma region.(2)For the detection of lung nodules,this paper designs a pulmonary nodule detection algorithm based on the YOLOv3 network,and optimizes the algorithm mainly from two aspects: First,the hybrid dilated convolution block is introduced to replace the convolution in the YOLOv3 feature extraction network.Batch normalize the activation layer,and stack the feature layer of the shallowest feature layer(52×52×256)output by the backbone feature network and the feature layer(104×104×128)immediately above it to improve the detection accuracy of the algorithm for small target lung nodules and reduce the missed detection rate.The second is to introduce a local Loss loss function based on cross entropy improvement for the problem of unbalanced positive and negative samples in the training process,which can correct the update direction of the model gradient.Finally,in the experimental part,the comparative analysis is carried out for many times,and it is concluded that the detection accuracy of the pulmonary nodule detection network proposed in this paper has reached 95.02%,which is about 6.02% higher than that of the improved network.(3)Based on the pulmonary nodule detection algorithm of YOLOv3,this paper designs an auxiliary pulmonary nodule detection system,which realizes the automation of pulmonary nodule detection to a certain extent and helps doctors reduce the heavy workload.This paper takes lung nodules in CT images as the research object,successively completes the CT image segmentation and nodule detection work,and designs an automated auxiliary detection system,which has good application value.