Pulmonary Nodule Detection Based On Deep Learning

At present,with the sharp increase in the number of lung cancer cases and deaths,people are paying more and more attention to the results of lung examinations.Pulmonary nodules are an obvious symptom in the early stage of lung cancer.Although lung CT images can be used for screening,the selection of the current algorithm has become the main limiting factor in the screening efficiency of pulmonary nodules.Therefore,this paper chooses the YOLOX algorithm to carry out research on the detection of pulmonary nodules,and analyzes and studies the data obtained in the experiment,and then improves and optimizes the YOLOX algorithm,which can finally effectively solve the problems existing in the detection of pulmonary nodules.,so as to help doctors reduce their workload and reduce people’s morbidity and mortality.The main research contents of this paper are as follows:①Firstly,it analyzes the current research status of pulmonary nodule detection at home and abroad,then analyzes and studies the characteristics of pulmonary nodules,and compares and studies the existing target detection algorithms.Finally,the YOLOX algorithm is selected,and the YOLOX algorithm is studied and described in detail.②By downloading and analyzing the relevant files in the LIDC-IDRI database,1186 image datasets containing only pulmonary nodules were obtained through preliminary screening.Since other tissues other than the lung parenchyma will have various effects on training,this paper finally chooses to extract the lung parenchyma based on the OTSU optimal global threshold method.Then,on the extracted lung parenchyma images,lung nodules were annotated and data augmented to obtain a complete experimental data set.Finally,the detection experiment of pulmonary nodules based on YOLOX algorithm was carried out.The AP(mAP)value of YOLOX algorithm for detecting pulmonary nodules reached 88.22%.Its detection speed can reach 58.3FPS on GPU 3090,but problems such as missed detection and low detection accuracy will occur for small lung nodules.③In view of the problems of missed detection and low detection accuracy in the detection of small pulmonary nodules by the original YOLOX algorithm,the YOLOX algorithm was improved and optimized.The improvements include:first,adding shallow output in the CSP Darknet network,and then inputting it into PAFPN for multi-scale feature fusion to increase the scale of detection of smaller lung nodules.Then,the CA channel attention mechanism module is added between the output of CSP Darknet and the input of PAFPN to assist PAFPN to "pay attention" to the channel information containing lung nodules.Optimizing the loss function:Since it is found that Loss is not easy to converge and the convergence value is not very low in training and detection;therefore,IOU loss and BCEWithLogits loss are replaced by α-EIOU loss and VariFocal loss,respectively.Finally,the improved and optimized YOLOX algorithm can detect pulmonary nodules with an AP(mAP)value of 91.43%and a detection speed of 57.2FPS.Although the detection speed is reduced by 1.1FPS compared to the original YOLOX algorithm,its AP(mAP)value is increased by 3.21%,and the phenomenon of missed detection of small pulmonary nodules and low detection accuracy has been significantly improved.