Research On Medical Image Cancer Detection Method Based On Deep Learning

Cancer is the leading threats to human health.Therefore,early screening and accurate diagnosis of cancer are important.Medical imaging,as the main cancer diagnosis method,is widely used in clinical diagnosis.Increasingly refined medical images provide a wealth of useful information that plays a vital role in assisting doctors in making accurate diagnoses.However,the growing medical imaging data poses a challenge to the diagnostic efficiency of doctors.From the perspective of computeraided diagnosis,this paper studies the deep learning detection model suitable for cancer images,and assists doctors to improve the diagnostic efficiency.The main content of this thesis includes:1)A method based on Faster-RCNN for cancer imaging was proposed,using FasterRCNN as a detection model for medical images.From a data point of view,data augmentation methods are used to amplify samples to solve the problem of less labeled data in cancer images.For the pathological features such as cancer nodules and inconspicuous contours,the laplacian convolution layer was added to sharpen the nodule edges and highlight the nodule contour.The size and proportion of the anchor frame generated by the RPN(Region Proposal Network)in the FasterRCNN are modified for the small size of the nodule detection target.Experiments show that the method achieves the auxiliary diagnosis effect on the LIDC-IDRI data set,which greatly improves the diagnosis efficiency.2)A transposed convolution inception block was proposed.The block mainly uses multi-scale transposed convolution layer for feature extraction,1*1 convolution layer for channel compression and batch normalization layer for accelerated convergence.The structure of the network block is optimized in combination with the inception structure.In response to the problem that Faster-RCNN has a poor detection performance on small targets,the transposed convolution inception block and VGG16 have been used to improve Faster-RCNN.The detection performance is improved by improving the feature extraction performance of the backbone network in the Faster-RCNN on the small target nodules in the image space.Experiments show that the Faster-RCNN model combined with VGG16 and transposed convolution inception block has significantly improved detection of regularly sized nodules and small target nodules with almost no increase in detection time.Compared with the Faster-RCNN using Res Net101,the detection speed is faster and the detection of small targets is better.