Improved Lung Nodule Detection Algorithm Based On YOLOV5

The electron computed tomography images are grayscale images,which make the characterization information of pulmonary nodules in the images limited,and the morphology of pulmonary nodules is diverse,mainly irregular and round-like,and the size varies,which is not conducive to the feature extraction of the detection network,thus leading to the problems of high false detection rate and imprecise localization in the lung nodule detection network model.To address this problem,this paper proposes the YOLO-Transformer(YOLO-TF)target detection network using the YOLOv5 target detection network and Vision Transformer model,and further streamlines the parameters on the basis of the YOLO-TF network to design the YOLO-Transformer Lite(YOLO-TFLite)network.Currently,various types of network structures exist in the field of target detection.Before carrying out network innovation optimization,in order to select the network suitable for the lung nodule detection task,the selection experiments of the underlying algorithm are conducted in this paper.By comparing the best performing detection networks among the current two-stage detection networks and single-stage detection networks,a comprehensive comparison and analysis of the detection accuracy and detection speed of each network was performed,and the YOLOv5 network was finally selected as the baseline network by combining the actual application requirements and the test data of each network.To improve the detection accuracy of the network,this paper optimizes the design of YOLO-TF network by adding the Vision Transformer model to the YOLOv5 network,and proposes a cascaded parallel Transformer module for the extraction of global features and a Deform convolution module for the extraction of detailed local features.In addition,a dual-dimensional codec feature extraction network is designed,which has richer feature information than the original feature extraction network,and a feature overlay layer is used to merge local features with global features to obtain a feature map with multidimensional feature information.In addition,the feature reorganization network is further designed to perform secondary feature processing on the output of the feature extraction network,and the amount of feature information in the feature map is balanced by reorganizing and regrouping the feature maps of different network levels.Therefore,the YOLO-TF network can extract richer feature information compared with YOLOv5,and the feature reorganization network can improve the detection accuracy of detectors at each scale by second weighting the features,increasing the weights of effective features,removing redundant background features,and synthesizing the feature information at each network level.In order to optimize the parameter design of the network and reduce the parameter complexity and computational complexity,this paper lightens YOLOv5 by adopting a lightweight convolutional structure and reparameterization design.The YOLO-TFLite network is proposed to further streamline the module parameters,and the DGhost lightweight convolution module is designed,and the DGhost convolution module is reparameterized;this paper modifies the structure of the YOLO-TF feature reorganization network by outputting only single-scale features for the detection network after its aggregation of multidimensional features and reorganization operations;for the lightweighting of the detection network using the Anchor Free mechanism is used to generate circular detection frames.In this link,Non Maximum Suppression(NMS)calculation is eliminated,and the redundancy of the number of detection frames can also be reduced by judging the region of the target through the center point.Compared with YOLOv5 and YOLO-TF networks,the YOLO-TFLite network is mainly built with ultra-lightweight network modules and uses only a single Anchor Free mechanism detector for the target detection task,all these optimizations make the number of network parameters of YOLO-TFLite much lower than YOLOv5 and YOLO-TF.In the performance evaluation experiments of YOLO-TF and YOLO-TFLite,the LUNA16 dataset with 888 lung CT images was used in this paper,and all images were randomly divided into training and test sets according to the ratio of 8:2.Ablation experiments and comparison experiments were conducted in the paper,and the experimental results showed that the mean detection accuracy of YOLO-TF reached 0.918,which was improved by 0.056 compared with YOLOv5;the number of network parameters of YOLO-TFLite was 6.6 MB,which was only one-seventh of YOLOv5.The average detection accuracy of YOLO-TFLite is only 0.011 lower than that of YOLOv5,which still maintains a better detection accuracy.In order to better apply the designed models,a lung nodule detection system based on the YOLO-TF and YOLO-TFLite models is designed in this paper,which enhances the ease of use of the two models through a visual interactive interface and adds the functions of test report generation,export,and model update.