Research On Medical Image Pathology Detection Based On Deep Learning

Lung cancer is one of the most malignant cancers in the world and has a high incidence and mortality rate worldwide,which has a significant impact on human health.Studies have shown that detecting and treating lung cancer at different stages of its development can greatly help patients’ survival rate,so timely detection and treatment in its early stages is particularly important.Early stage lung canc er can be seen in the form of lung nodules,but in clinical practice,reliance on the naked eye to identify lung CT scans can lead to misdiagnosis and missed diagnoses.With the advent of computer-aided diagnosis(CAD)systems,doctors are able to assist i n medical diagnosis,but there are shortcomings in detection accuracy.Deep learning has been widely used in medical imaging,and has a good performance in terms of detection accuracy,and has obvious advantages in dealing with misdiagnosis and omission.Therefore,this paper uses deep learning methods to conduct an in-depth study on lung nodule detection and false positive rejection,and the main research content is as follows.(1)This paper firstly pre-processes the input CT images and segmentation of lung parenchyma.The CT image information is read and saved in PNG format through a series of operations such as pixel mapping and data type conversion,and the world coordinate information is converted into image coordinate information;then the lung parenchyma is segmented using binarization,pixel adjustment and image morphology operations;finally,the data set is integrated and saved in Pascal VOC format.(2)In the lung nodule detection stage,an improved YOLOv3 lung nodule detection algorithm is proposed in this paper.For the characteristics of "small" lung nodules,this paper incorporates the features extracted by the shallow network into the multi-scale feature prediction,and adds an improved hybrid attention module to the deep path,which can make the network focus on the features of small target lung nodules and ignore the irrelevant information;for the problem of performance degradation due to the deepening of the network model,this paper designs for the problem of performance degradation due to deeper network model,this paper designs a loopback residual module,which combines the jump link and residual feedback in the residual network to improve the DBL Block,effectively solving the network performance degradation problem;for the case of lung nodules with variable location and adhesion to other lung tissues,this paper selects CIo U for position loss calculation,effectively improving the precise location of lu ng nodules.(3)In the false positive rejection stage,this paper investigates the 3D CNN algorithm C3 D,and selects Focal Loss as the loss function to balance the focus on positive and negative samples in training,effectively reducing the false positive rate of candidate nodules obtained by the YOLOv3 algorithm and improving the accuracy of the lung nodule detection task.(4)This paper verifies that the improved YOLOv3 algorithm and the C3 D algorithm are more accurate for lung nodule detection using the dataset LUNA16.Experiments using three and four prediction scales,experiments usin g different attention mechanisms and improved hybrid attention,ablation experiments with improved design ideas,and experiments using other target detection algorithms an d the improved YOLOv3 validated the improved YOLOv3 lung nodule detection algorithm for higher detection accuracy;the analysis using ablation experiments validated the false positive lung nodule detection accuracy after data enhancement and Focal Loss calculations resulted in lower false positive rates and higher accuracy of lung nodule detection.