Lung CT Imaging And Analysis Algorithm Research Using Deep Learning

In recent years,thanks to the advantages of real-time,non-invasive imaging and flexible operation,lung X-ray computer-assisted tomography(CT)equipment has gained great popularity in hospitals at all levels at home and abroad,and has been widely used in the screening of many lung diseases,such as early screening of lung cancer and Rapid screening for infectious pneumonia.Lung cancer is the leading cause of cancer deaths,and the purpose of early lung cancer screening is to identify patients with asymptomatic early-stage disease so that early treatment can be provided,which can prevent malignant transformation of the lesion,facilitate clinicians to tailor a reasonable diagnostic plan and provide targeted control,effectively curb the spread of cancer cells and improve the cure rate.In addition,rapid screening of infectious pneumonia based on lung CT can effectively control the spread of infectious infection sources and plays an important role in the safe prevention and control of epidemics.However,lung CT-based disease screening can place a high workload on pulmonary clinicians.On the one hand,there are a large number of high-incidence lung diseases,especially infectious pneumonia,and a large number of positive patients break out in a short period of time,thus bringing clinicians a large amount of screening work;on the other hand,tiny lesions and blood vessels have similar ranges of CT values during lung CT scanning,resulting in target lesions(potential malignant tumors in lung cancer screening,ground glass nodules in infectious pneumonia screening,and other lesions)are highly likely to be flooded in the imaging area,posing great screening difficulties for clinicians.Nevertheless,CT-based lung disease screening still faces many challenges and difficulties.CT imaging and diagnosis are crucial in lung CT disease screening.On the one hand,CT imaging of the lung provides accurate 3D images of internal structures through differences in X-ray absorption by lung tissues and organs,and low-dose imaging can avoid radiation damage to patients and equipment paralysis due to screening of large crowds.How to reduce radiation dose while ensuring imaging quality has been a key issue in the task of lung CT screening.On the other hand,lung CT diagnosis provides clinicians with lesion detection through computeraided algorithms,but because normal capillaries and lung lesions have a similar range of CT values and present a structure similar in shape and size to the lesion in the imaging field,it can easily lead to a large number of false-positive detection samples,and the same lesion has different morphology at different developmental periods,leading to certain intraclass differences in the process of lesion detection.The same lesion has different morphology at different stages of development,leading to certain class differences in the process of lesion detection and increasing the difficulty of lung disease screening.Deep learning-based approach can effectively improve the diagnostic efficiency of lung CT disease screening and effectively address the key issues arising in the imaging and diagnostic process.On the one hand,high-performance deep learning imaging algorithms are designed to maintain the quality of lung imaging while reducing radiation dose to meet clinical needs.On the other hand,facing a large amount of lung CT disease screening work,designing high-quality lung CT diagnosis algorithms based on deep learning models can effectively improve lesion detection accuracy,assist clinicians in intelligent screening,avoid misdiagnosis and omission caused by human factors,and eliminate potential safety risks.To this end,this paper is based on deep learning to study lung CT imaging methods for lung CT disease screening tasks to solve the problem of low-dose lung CT imaging quality;taking lung nodule false positive rejection and new crown detection tasks as examples,we study intelligent diagnosis related methods for lung CT screening to improve the efficiency of lung CT lesion screening and reduce the computational complexity of lung CT data.The main research of this paper includes deep neural network-based lung low-dose lung CT imaging,automatic implementation of lung nodule false-positive rejection and new coronavirus detection,and the main work and contributions are as follows:(1)To reduce the radiation dose and improve the CT image quality,a new dual scale residual attention weighted network(DRAWNet)is proposed for improving low-dose CT imaging quality.In order to enrich the context information,we introduce a dilated convolution to expand the acceptance domain of convolution neural networks(CNNs)and effectively reduce computation complexities.For integrating different scales of spatial features and enhance the spatial features,two-scale attention mechanism is designed to integrate the features of different dimensions and remove irrelevant information.The proposed network consists of three interactive functional components: the basic module(BM)is used as the infrastructure of the whole network to extract spatial features,and the two-stream attention module(TSAM)is used to maintain high resolution and extract more extensive contextual features.Feature enhancement module(FEM)is used to enhance depth features and avoid overfitting.The proposed DRAWNet was evaluated on the AAPM Mayo Clinic LDCT grand challenge dataset and its performance was superior to the most advanced competitive methods.This efficient,accurate and reliable LDCT denoising method has great clinical application potential.(2)To improve the lung nodule detection efficiency,a multi-dimensional nodule detection network(MDNDNet)is proposed for automatic false positive nodule reduction.This paper cooperatively integrates multi-dimensional nodule information,designs threedimensional CNN to extract comprehensively and complementary volume correlation features of lung nodules,and designs two-dimensional CNNs with attention module to extract 2D spatial features from sagittal plane,coronal plane and transverse plane.The proposed MD-NDNet consists of three interdependent functional parts: the preprocessing module is used to initialize the original candidate nodules and solve the problem of data imbalance.The multi-stream network(MSN)is used to integrate multi-dimensional nodule information and extract comprehensive spatial features and inter-plane volume correlation features,and the multi-scale Network(MSN)is used to cover different sizes of nodule candidates and reduce intra-nodule differences and inter-class similarities.This method has been evaluated on the ten-fold cross validation on the LUNA-16 Challenge Dataset of ISBI 2016.Experimental results show that the proposed method achieves good classification performance in the term of the CPM score of0.9008.All these results show that our method can effectively,accurately and reliably detect lung nodules for clinical diagnosis.(3)To enhance the COVID-19 detection efficiency,the COVIDNet is proposed for automatic and accurate patient-level COVID-19 detection.The three-dimensional(3D)multi-scale network(3D MSN)is designed to extract multi-dimensional inter-plane correlation features of typical covid-19 lesions(GGOs).In order to cover more features of GGOs lesions and reduce intra-class differences,phase ensemble(PE)module is proposed to aggregate different phases in CT scanning.The proposed method is evaluated on COVID-19 dataset and we employe the five-fold cross validation.Experimental results show that this method achieves classification performance with specificity of 0.9900,sensitivity of 0.9300,accuracy of 0.9600 and accuracy of 0.9894.All these show that our proposed method can effectively,accurately and reliably detect whether patients have covid-19 in the diagnosis of COVID-19.In summary,the thesis focuses on the research of lung CT imaging and analysis algorithms,based on deep learning,from lung CT imaging to intelligent assisted analysis,lung CT imaging using DRAWNet for low-dose CT denoising in the image domain to improve imaging quality,lung CT intelligent assisted analysis for lung nodule false-positive reduction and Covid-19 detection,proposing MD-NDNet and COVIDNet to improve the efficiency of pulmonary clinical diagnosis.The above-mentioned works can effectively alleviate social tension between doctors and patients and reduce the tension of medical resources under the new crown epidemic.It is of great significance to promote public health security and cyberspace security.