CT Image Automatic Detection System Based On Deep Learning

Lung nodules are key clues in the detection of early-stage lung cancer.Traditional pulmonary nodule detection is often completed based on algorithms such as threshold,clustering,or template matching.However,due to the relatively small size of pulmonary nodules and inconspicuous features,traditional algorithms are often cumbersome to implement and have limited accuracy.Recently with the rapid development of computer vision,deep learning-based pulmonary nodule detection methods no longer require artificially designed features,and the accuracy has been greatly improved,gradually becoming a preferred choice for people.In this thesis,an automatic detection algorithm of pulmonary nodules in lung CT images is implemented based on deep learning.The automatic detection algorithm of pulmonary nodules can be mainly divided into image preprocessing and a three-stage deep learning algorithm.By performing image enhancement and lung parenchyma segmentation on lung CT images,the accuracy of subsequent processing can be improved.The three-stage algorithm for automatic detection of pulmonary nodules is as follows:1)Lung nodule candidate region extraction algorithm,which consists of two different models: a model based on Faster R-CNN that introduces FPN and improves the anchor design method;based on U-Net,an SE module is introduced and added Model with residual connections.The outputs of the two improved models are combined in the prediction stage to improve sensitivity.2)False-positive nodule suppression algorithm,which is based on a deep convolutional neural network and adopts three feature extraction enhancement modules:Res Ne Xt,Dense Net,and Res2 Net.Finally,a network-weighted voting method loaded with three modules is used to make predictions.3)Pulmonary nodule fine segmentation algorithm,which takes UNet++ as the main body,introduces the Res2 Net module at the encoding end to enhance feature extraction,and embeds the sc SE module at the decoding end to recover the nodule features by combining spatial and channel information.In the prediction stage of the model,the three-stage algorithm is integrated into a series to achieve end-to-end prediction.The model achieves 94.1% sensitivity and83.61% segmentation Dice coefficient on the public LIDC-IDRI dataset,verifying the effectiveness of the algorithm.At the same time,based on the Vue and Spring boot frameworks,this thesis also deploys the proposed automatic detection model of pulmonary nodules to the WEB site,which response to the policy call of "Internet + medical health",realizes remote online access and use,and makes medical resources Deficient areas can also enjoy convenient online CAD for the benefit of early-stage lung cancer patients.