Research Of Lesion Detection With Deep Learning Algorithm In Medical Image

With the rapid development of medical imaging technology,health care has entered a new era of big data.Among these,90% of medical big data come from medical imaging.Therefore,it is a very meaningful study to extract valuable information from massive medical image data quickly,which is effective for clinical prevention,diagnosis,treatment and medical research.The lesion detection task is analyzed and studied through various modes of medical images to achieve the location of lesion,suspected lesion,and the type of disease.The lesion object detection algorithm based on deep learning has the characteristics of high accuracy,fast speed and strong generalization.Especially,deep convolutional neural network can extract the lesion features in different patterns effectively.Therefore,the study of lesion detection algorithms based on deep convolutional neural networks have important theoretical and industrial value.The main work of the thesis includes the following two aspects*:(1)In order to improve the lesion detection accuracy of computer-aided diagnosis of in chest X-ray images,chest X-ray lesion detection algorithm(DDFCN)based on deep deformable full convolutional networks(FCN)is proposed in our paper.Aiming at problems of the large amount image data of chest X-ray,multiple lesions with poor reproducibility,inconspicuous features of some lesions,and time-consuming and laborious manual analysis and strong subjectivity.Due to the limitations of conventional convolution rectangular sampling,the features of lesion obtained through conventional convolution network cannot reflect themselves well,which will lead to low detection accuracy of the network model.Therefore,the deformable convolution is used in the paper to enhance the ability of the network to extract features.Firstly,Residual Networks(Res Net)is used to construct the lesion feature extraction sub-network,and the partial convolution layer in the feature subnet is improved by deformable convolution to enhance the lesion representation for feature extraction sub-network in chest X-ray images.Then,using the feature map of the entire chest X-ray image obtained the feature extraction sub-network enhances scores of classification response and regression response through the position-sensitive region of Interest(Ro I)pooled network layer established by the deformable convolution model.Finally,though Softmax performs classification and position regression,the acquired lesion informations are subjected to post-processing network layer,such as Soft-NMS,to achieve lesion detection in chest X-ray images.In comparative experiment,the single lesion detection tasks and multiple lesion detection tasks in chest X-ray images are used to verify the high accuracy of detection of lesions in chest Xray images.(2)In order to improve the accuracy of lesion detection in small medical image datasets,we proposed a breast cancer lesion detection algorithm based on feature pyramid networks with deformable convolution and transfer learning.Aiming at solving the problem that is difficult to obtain for medical image acquisition and high quality labeling,as well as various shapes and scales of the lesion,the paper solves the above problems using transfer learning combined with the deep deformable convolution feature pyramid network.Firstly,because of the the suggesting next dimension on residual networks(Res Ne Xt),the impact of FPN on the computational complexity of the model can be compensated to some extent.So,we reconstruct some convolutional layers of Res Ne Xt to the improved deformable convolutional layers,and combining the semantic features of some high-level networks with the previous layer features to establish the feature extraction sub-network,which is based on the Feature Pyramid Networks.Secondly,the multi-layer fusion of FPN leads to the inter-layer effect,and we can use a convolution operation to solve this problem.The feature maps of each layer of FPN are obtained,and then we use the Region Proposal Networks(RPN)extracts the regions of interest on these features.Thirdly,in order to obtain better pooling feature maps,an improved deformable convolution model is used to establish the region of interest pooling network,and then we use some post-processing network layers to obtain breast cancer lesion location and classification.Finally,we use the deep transfer learning,TDFNet is trained on the Deep Lesion datasets with large data.Then,the feature extraction sub-network layer in TDFNet is frozen,and the finetuning the model on the DDSM and MIAS datasets.It improve the accuracy of TDFNet detection of breast cancer lesions on small MIAS datasets.The comparison experimental results show that the proposed algorithm has higher detection accuracy for lesion on small sample datasets.