Establishment And Application Of Multiple Parameters Model Of Lung Cancer Screening Based Low-Dose Spiral CT,Tumor-Related Antibody Panel And Deep Learning

Background and objectiveWith the increasing urbanization and the acceleration of aging process in China,lung cancer is one of the important diseases that endanger human health,and its morbidity and mortality are increasing yearly.In 2017,the status and trends of Chinese tumors showed that tumors were the leading cause of death for both urban and rural residents of China.The incidence rate of lung cancer in China is the second highest among females and the highest among males.The mortality rate of lung cancer in China is the highest among both males and females,which accounts for21.68%of the seven malignant tumors before death.R.Peto,the British oncologist,has predicted that if the smoking and air pollution fails to be controlled timely in China,there will be more than one million new cases of lung cancer by 2025.Early diagnosis and treatment of lung cancer through screening has been recognized as the most effective way to prevent and control lung cancer and is an important measure to reduce mortality of lung cancer.Since Brenner firstly proposed the concept of low-dose computed tomography（LDCT）in the early 1890’s,many scientists around the world have studied the application of LDCT in lung cancer screening.The results of the NLST showed that 20 percent reduction in lung cancer mortality by LDCT of screening in high-risk groups was informed compared to X-ray of chest.The international early lung cancer action plan（I-ELCAP）,Dutch Belgian randomized lung cancer multi-slice CT screening trial（NELSON trial）and the national lung screening trial（NLST）are the three largest LDCT lung screening programs in the world.They have separately introduced guidelines for lung cancer screening that conform to the characteristics of each region and recommended LDCT lung cancer screening in high-risk groups and Lung-RADS classification management for pulmonary nodules.Due to its own epidemiological characteristics of lung cancer in China,we can’t completely copy western developed countries LDCT lung cancer screening program.In order to achieve the requirement of early diagnosis and treatment of lung cancer,and to avoid excessive diagnosis or treatment,we should emphasize on the projects on early detection and treated of the malignant tumor in high-risk group,or health economics research,and on positive efforts to get more data for further study.We should also establish suitable LDCT lung cancer screening for basic conditions of urban in China and efficient economic system of screening.The contents of this study are mainly from the following four aspects1.Based on the data of LDCT Lung cancer screening conducted by Henan tumor hospital from 2013 to 2018,the epidemiological characteristics of Lung cancer in cities in central China were summarized and the application of Lung-RADS classification in Lung cancer screening was valued.2.Due to the deficiency of lung-RADS 1.1 in Lung cancer screening in Chinese,the risk stratification of Lung-RADS 1.1 in ground glass nodules is modified to provide methodological support for the early detection of lung cancer.3.In order to establish a high-quality and efficient screening models for lung cancer,a CAD system based on 3D deep convolutional neural network（3D-DCNN）,combined with the visual read by the radiologists,was used to evaluate the LDCT image data of the volunteers who are screened for early lung cancer in our hospital.4.A multi-parameters model of lung cancer screening based on deep learning and tumor-related antibody spectrum was established to explore a set of high quality and efficient lung cancer screening and diagnosis system suitable for Chinese.The main innovation points of this study1.In order to reduce the rate of missed diagnosis or false positive,and to avoid unnecessary follow-up with LDCT or treatment excessively,the Lung-RADS classification and the CAD system were introduced to improve the standardized management of pulmonary nodules in early lung cancer screening.2.The complementary Lung-RADS 1.1 was used to manage the risk stratification of ground-glass nodules for improving the detection rate of early lung cancer.3.The detection rate of pulmonary nodules was improved by using CAD system based on the 3D-DCNN combined with visual reading of radiologists.4.The recontruction of multiple parameters model of lung cancer screening based on deep learning and 7-TAAbs is better for LDCT lung cancer screening.Part 1 The epidemiological characters of lung nodules detected by LDCT and its risk grading management by Lung-RADS systemObjectivesTo analyze the non-calcified pulmonary nodules and lung cancers by low-dose computed tomography（LDCT）screening and the relevant epidemiological factors,as well as to explore the age threshold of lung cancer screening and the clinical application of Lung-RADS classification.Materials and methodsThe 7683 cases with high-risk groups and 1167 cases of volunteers with physical examination were selected during the same period in our hospital.1339 cases were divided into high-risk groups according to the severity of smoking or second-hand smoke（smoking history is more than 30 years,smoking cessation is less than 15 years,or the history of second-hand smoking history is greater than 20 years）,6344 cases were grouped as middle-risk group（those with or without smoking history or second hand smoking history who could not meet the smoking groups conditions）;exposure recorded were divided into ever smokers（smokers or light smokers）and the 1167cases of health screening were grouped as the low-risk group.According to the gender,there were 4900 males and 3950 females,and there were 110 cases in less than 40year old group,794 cases in 40-45 years old group,1633 cases in 45-50 years old group,2505 cases in 50-55 years old group and 3858 cases in Greater than or equal to55group according to age group.The I-ELCAP classification scheme（the positive nodule was defined as noncalcification,solid or part of solid nodules with at least a diameter of 5mm or noncalcification,nonsolid nodule with diameter 8mm at least）and the Lung-RADS grading（Lung-RADS 3 or 4 was positive）were used for the classification of lung nodules.The classification of nodules and the distribution of pulmonary segments were statistically analyzed.Lung cancer staging was staged according to the 8^thedition of TASLC.ResultsAll the 8850 cases were screened by the low dose CT,the positive nodules were860 cases and the positive rate was 9.72%（860/8850）according to the I-ELCAP classification system,the positive rate of lung nodules based on the Lung-RADS classification was 6.67%（590/8850）,and the positive predictive rate of lung cancer was 2.54%（15/590）,all missed cases of lung cancer were presented as pGGNs on CT.Among the 17 cases of lung cancer patients,there were 11 cases with super early lung adenocarcinoma（0,ⅠA andⅠB）accounting for 64.7%（11/17）,in which there were5 males and 12 females,and the difference was statistically significant（12/3950:5/4900,χ²=4.618,P=0.031）.According to the Lung-RADS classification,the positive rate of lung nodules between male and female was statistically significant（7.5%vs5.6%,P=0.00）.The positive rate of pulmonary nodules and the nodule classification was gradually increased with gender（all P<0.05）,especially in the 40-45 year-old group.The positive rates of lung nodules between high-risk groups,middle-risk group and low-risk group were 10.75%（144/1339）,6.76%（429/6344）and 1.46%（17/1167）separately,and the incidence rate of lung cancer in high-risk group were 0.53%in high-risk group,0.16%in middle-risk group（10/6344）,and 0.17%in low-risk group,the positive rate of nodules and the incidence of lung cancer in high-risk group were obviously higher than that of middle-risk and low-risk group（χ²=467.386,7.337,P=0.00,0.026）.ConclusionsThe Lung-RADS classification can improve the positive predictive rate of lung cancer or,reducethe diagnostic accuracy of lung noduless the false positive rate combing with the changes feature of nodules during follow-ups in lung cancer screening.The standard of setting the 40 years old as the threshold of screening is more suitable for China’s national conditions.Part 2 Effectiveness and feasibility of complementary lung-RADS version 1.1 in risk stratification for pGGNs in LDCT lung cancer screening in Chinese populationObjectivesTo evaluate the effectiveness of using a modified lung imaging reporting and data system 1.1（Lung-RADS 1.1）for risk stratification of pure ground-glass nodules（pGGNs）in low-dose computed tomography（LDCT）for lung cancer（LC）screenings in China.Materials and MethodsEight subjects with nine pGGNs originated from Cancer Screening Program were enrolled as training set and 32 asymptomatic subjects with 35 pGGNs were selected as validation set from November 2013 to Octomber 2018.The complementary Lung-RADS categories were set based on the GGN–vessel relationship（GVR）.The correlations between GGN-vessel relationships and pathology were evaluated,and the diagnostic value of complementary Lung-RADS version 1.1 based on GVR in discriminating malignant pGGNs were analyzed.ResultsThe inter-reader agreements for Lung-RADS 1.1（intraclass correlation coefficient（ICC）=0.999）and complementary Lung-RADS 1.1（ICC=0.971）displayed good reliability.The combined incidence of invasive adenocarcinoma in type III and IV was more than that of benign and preinvasive diseases（30%vs 75%,P=0.013）.Type II GVR between two benign（66.7%）,seven preinvasive（53.8%）,and six invasive（21.4%）GGN cases was statistically significant（χ²=5.415,P=0.019）.GGN pathological groups and GVR had a significant correlation（r=0.584,P=0.00）.Compared to Lung-RADS 1.1,complementary Lung-RADS 1.1 had better performance in the training set,with its sensitivity increased from 33.3%to 88.9%,accuracy increased from 44.4%to 88.9%,false-negative proportion（FNP）decreased from 66.7%to 11.1%,and the sensitivity to predict malignant nodules increased from13.8%to 93.1%,accuracy increased from 28.6%to 80.0%,and FNP decreased from86.2%to 6.9%in validation set.The detection rate of preinvasive disease and adenocarcinoma was increased from 12.5%to 90.6%and that of missed diagnosis decreased from 87.5%to 9.4%in the validation set,P=0.004.ConclusionComplementary Lung-RADS 1.1 is superior to Lung-RADS 1.1 and benefits for LC screening of LDCT in China.Part 3 Application of computer aided diagnosis system based on 3D deep convolutional neural network in lung cancer screeningObjectivesTo evaluate the value of CAD system based on 3D-DCNN combined with visual detection（VD）in the diagnosis of lung nodules and lung cancer in LDCT lung cancer screening,and to select a set of high–quality and efficient lung cancer screening methods.Materials and MethodsThe LDCT imaging data of 8850 lung cancer screening volunteers with 1111nodules of lung and the clinical information of patients diagnosed with lung cancer were retrospectively analyzed from November of 2013 to December of 2018.The routine LDCT was examined for all volunteers at least once in this study.There are three ways to read the imaging:VD method,CAD method,Combination between the VD and CAD method.The diagnostic criteria of pulmonary nodules were observed by two senior radiologists with chest imaging,and the consistent opinion was taken as the true nodules,and the cumulative number of true nodules detected by CAD system was taken as the reference standard.In terms of the number of nodules,type of nodules and the classification of Lung-RADS of nodules,the missed diagnosis rate,the detection rate and the number of false positive rate of nodules were compared in percentage（%）withχ²test between groups.Lung cancer staging system of version 8^threvised by TASLC was used in our study.ResultsA total of 8850 cases of volunteers were analyzed by two senior radiologists of chest radiology and 1111 true nodules were detected,and there were 590 nodules with diameter more than 6mm and 521 nodules with diameter of less than 6mm,which were taken as the reference standard.Compared with the VD method,the detection rate of nodules in the CAD method and the VD combined with CAD method was significantly increased,and the rate of missed diagnosis was significantly reduced（80.1%vs 94.2%,95.7%;19.9%vs 5.8%,4.3%,χ²=101.65,128.5,P=0.00,0.00）.Compared with CAD method,the detection rate of nodules in VD combined with CAD method was higher,and the rate of missed diagnosis was lower（χ²=10.38,P=0.006）.The number of false positive nodules detected by CAD method was higher than those other two methods.The detection rate of nodules with Lung-RADS 2 was the lowest in the VD method,which was 60.3%(314/521.However,combined with CAD method,the detection rate of nodules with Lung-RADS 2 was significantly improved to 94.5%（490/521）.Compared with the VD method,the detection rate of Lung-RADS classified nodules by CAD method and VD combined with CAD method was significantly increased（χ²=25.083,23.449,P=0.000,0.000）,while the detection rate of Lung-RADS classified nodules between the CAD method and the VD combined with CAD method was not statistically significant（χ²=2.086,P=0.072）.Compared with the VD method,the detection rate of different types of nodules by CAD method or the VD combined with CAD method was significantly increased,and the difference was statistically significant（χ²=6.955,6.821,P=0.031,0.033）.However,the difference between CAD method and the VD combined with CAD method was not statistically significant,P>0.05.The difference of positive prediction rate,missed diagnosis rate and false positive rate between the VD method and VD combined with the CAD method were not statistically significant（χ²=0.006,P=1.000）.Compared with the VD method and the VD combined with the CAD method,the positive prediction rate of lung cancer by CAD method was significantly reduced,the rate of missed diagnosis and the false positive rate were significantly increased,and the difference were statistically significant（88.2%vs 94.1%,94.1%;9.2%vs 2.1%,2.1%χ²=14.605,14.693,P=0.002,0.002）.ConclusionThe CAD based on 3D-DCNN combined with VD can improve the detection rate of true pulmonary nodules and reduce the detection of false positive nodules,which is the preferred method using in the large scale LDCT lung cancer screening for high-risk population.Part 4 Establishment and value of multiple parameters model of lung cancer screening based tumor-related autoantibody panel and deep learningObjectivesTo establish the multiple parameters model of lung cancer screening based deep learning and seven tumor-related autoantibodies（7-TAAbs）and to evaluated its value in clinical.Materials and MethodsThere were 196 patients with 201 observations enrolled in this retrospective study,78 lesions were proved by follow-up and 123 lesions were proved by pathology.the data of these patients were randomly divided into training sets（146 patients,151observations）and validation sets（50 patients,50 observations）.In particular,196patients with suspected LC based on CT images and a panel of TAAbs against seven TAAs（p53,PGP9.5,SOX2,GAGE7,GBU4-5,CAGE and MAGEA1）detection were selected.Nodules with medium or high malignant risks defined by DL were defined as positive nodules,and those with positive expression of 7-TAAbs were defined as having elevated AAbs assay signals for any one of the seven antigens（p53,PGP9.5,SOX2,GAGE7,GBU4-5,CAGE and MAGEA1）in the TAAbs panel.The per-observation estimates of the diagnostic performance(sensitivity,specificity,accuracy,positive predictive value（PPV）,negative predictive value（NPV）,positive likelihood ratio（+LR）,Youden index of DL,and 7-TAAbs for lung carcinoma diagnosis were computed separately.In the discrimination of the malignant nodules,the performance was assessed via a receiver operating characteristics（ROC）curve analysis for both the DL and 7-TAAs approaches.ResultsFor the training set,the combined DL and 7-TAAbs approach yielded a higher specificity（98.5%compared with 87.7%and 84.6%,respectively）,PPV（97.4%compared with 91.2%,and 81.5%,respectively）,and+LR（28.67 compared with 7.85and 3.32,respectively）as compared to the individual DL and 7-TAAbs approaches,but a lower sensitivity（43.0%compared with 96.5%and 51.2%,respectively）,accuracy（66.9%compared with 92.7%and 65.6%,respectively）,and Youden index（0.42 compared with 0.84 and 0.36,respectively）as compared to the individual DL and 7-TAAbs approaches.In addition,for the validation set,the sensitivity of the DL approach in LC diagnosis is higher than that in the 7-TAAbs and combined DL and7-TAAbs approaches（90.0%compared with 40.0%and 30.0%,respectively）;however,the combined DL and 7-TAAbs approach yielded a higher specificity（95.0%compared with 90.0%and 65.0%,respectively）and PPV（90.0%compared with 79.4%and 85.7%,respectively）as compared to the individual DL and 7-TAAbs approaches.Compared with the DL or 7-TAAbs approach,the combined DL and7-TAAbs approach yielded a low NPV（47.5%compared with 81.3%and 50.0%,respectively）,accuracy（56.0%compared with 80.0%and 60.0%,respectively）and Youden index;accompanying with an increase in+LR was observed（6.0 compared with 2.57 and 4.0,respectively）.For the training set,the AUC for the DL approach was 0.908（95%CI:0.852–0.964）and for the 7-TAAbs approach,it was 0.692（95%CI:0.608–0.776）,while that for the combined DL and 7-TAAbs approach,it was0.708（95%CI:0.627–0.789）.Furthermore,for the validation set,the AUC for the DL approach was 0.750（95%CI:0.602–0.898）and for the 7-TAAbs approach,it was0.650（95%CI:0.498–0.802）;in comparison,the combined DL and 7-TAAbs approach yielded an AUC of 0.625（95%CI:0.471–0.779）.Compared with the individual DL and 7-TAAbs approaches,the DL or 7-TAAbs approach showed a remarkable increase in sensitivity（96.7%compared with 90.0%and 40.0%,respectively）and NPV（90.9%compared with 81.3%and 50.0%,respectively）for the validation set;however,for the combined approach,these values were lower than those of the DL approach on the training set（sensitivity:95.5%compared with 96.5%;NPV:92.3%compared 95.0%）,which is attributed to the low sensitivity of the 7-TAAbs approach.Furthermore,the DL showed a higher accuracy（80.0%compared with 60%and 78%,respectively）and Youden index（0.55compared with 0.3 and 0.47,respectively）than those of the 7-TAAbs,and DL or7-TAAbs approaches.Moreover,the 7-TAAbs approach showed the highest PPV and+LR values for both the training and validation sets;however,it had the highest specificity（90.0%compared with 65%and 50%,respectively）only for the validation set because of the randomization factor specified earlier.The AUCs for LC screening using the DL or 7-TAAbs approach were 0.842（95%CI:0.771–0.914）for the training set and 0.758（95%CI:0.609–0.907）for the validation set.Conclusion:The multiple parameters model of lung cancer screening based on deep learning and 7-TAAbs has a excellent performance in clicinal.