A Study Of Artificial Intelligence In The CT-based Diagnosis Of Early Lung Adenocarcinoma And COVID-19

Part one CT-based radiomics to evaluate the invasiveness of early lung adenocarcinoma manifesting as ground-glass nodulesBackground:With an increased awareness of the importance of physical examination,the widespread use of computed tomography(CT)has increased the incidental detection of ground-glass nodules(GGNs).The pathology of ground-glass nodules can be benign,preinvasive,or invasive lung adenocarcinoma.In 2011,the International Association for the Study of Lung Cancer,American Thoracic Society,and European Respiratory Society proposed a new classification system for lung adenocarcinoma.Lung adenocarcinomas are classified as atypical adenomatous hyperplasia(AAH),adenocarcinoma in situ(AIS),minimally invasive adenocarcinoma(MIA),and invasive adenocarcinoma(IAC).AAH and AIS were defined as preinvasive lesions.Among them,the 5-year recurrence-free survival rate of patients with preinvasive lesions and MIA is nearly 100%,which are defined as indolent adenocarcinomas and recommended as a follow-up or limited segmental/lobectomy.In comparison,patients with IAC generally used lobectomy with systemic lymph node dissection.Therefore,accurately distinguishing indolent adenocarcinoma from IAC before surgery is vital to guide clinical treatment strategies.Objective:This study aimed to establish a nomogram that integrates clinical characteristics,imaging morphology,and radiomics features to evaluate the invasiveness of early lung adenocarcinoma manifesting as GGNs.Method:We retrospectively analyzed 509 cases from January 2016 to October 2018 confirmed by pathology as early lung adenocarcinoma,manifesting as focal GGNs with a diameter?of 3cm on CT.Two radiologists with 12 and 7 years of experience independently evaluated ground-glass nodules' imaging morphology and further segmented on the IntelliSpace Discovery platform(ISD,Philips Healthcare,Best,The Netherlands).70%of cases were included in the training cohort and 30%in the validation cohort.The radiomics features were extracted by the Max-Relevance and Min-Redundancy(mRMR)and the least absolute shrinkage and selection operator(LASSO).We uesd the finally selected features to calculate the radiomics score.The invasiveness-related clinical and CT morphological predictors were determined by univariate and multivariate logistic regression.The nomogram was constructed by combining clinical characteristics,CT morphology,and radiomics score.The receiver-operating characteristic(ROC)curve was used to show the nomogram's discrimination performance.The calibration curve was plotted to assess its calibration,and the decision curve analysis(DCA)was conducted to determine the clinical utility of the nomogram.Results:(1)Age,smoking history,long diameter,and average CT value were retained as independent predictors for invasiveness of early lung adenocarcinoma manifesting as GGNs.(2)Mean radiomics score for invasive adenocarcinoma was higher than those for indolent adenocarcinoma(training set 1.53 vs-0.97,P<0.001;validation set 1.36 vs-1.03,P<0.001),with statistical difference.Radiomics score exhibits a favorable discrimination performance,with an AUC of 0.892(95%CI,0.860-0.926)in the training set and 0.892(95%Cl,0.838-0.947)in the validation set.(3)The nomogram,which incorporated age,smoking history,long diameter,average CT value,and the radiomics score,yielded an AUC of 0.940(95%CI,0.916-0.964)in the training set and 0.946(95%CI,0.907-0.986)in the validation set,better than that of single radiomics score(Delong's test,P<0.001).Conclusion:This nomogram can accurately predict the invasiveness of early lung adenocarcinoma manifesting as GGNs before surgery,providing a non-invasive predictive tool for clinicians to develop individualized treatment strategies.Innovation points:(1)Using the advanced radiomics method,explore the unrecognizable high-order features by human eyes on CT to predict the invasiveness of early lung adenocarcinoma manifesting as GGNs.(2)Using multi-dimensional data,including clinical characteristics,CT morphology,and radiomics features,to intuitively and accurately evaluate the invasiveness of early lung adenocarcinoma manifesting as GGNs,and to provide a non-invasive predictive tool for clinicians to develop individualized treatment strategies.Part two CT-based Deep Learning to Differentiate COVID-19 from Community-acquired pneumoniaBackground:The coronavirus disease 2019,named COVID-19 on February 11,2020,is an ongoing pandemic.The World Health Organization(WHO)recognized COVID-19 as a "public health emergency of international concern" on March 11,2020,making it a public health crisis globally.Although the reverse transcription-polymerase chain reaction(RT-PCR)test serves as the gold standard for diagnosing COVID-19 patients,the high false-negative rate of this assay has restricted the prompt diagnosis of infected patients.As a common diagnostic tool,CT is easy and fast to use and is complementary to RT-PCR.However,the similar CT appearances between COVID-19 and community-acquired pneumonia(CAP)result in misdiagnoses.Therefore,accurate diagnosis of COVID-19 without delay remains a crucial challenge for epidemic prevention and control.Objective:This study aimed to develop and test a deep learning-based automatic framework to differentiate COVID-19 from CAP.Additionally,it aimed to compare the performance of the model with the findings from clinicians,and explore the auxiliary diagnostic role of artificial intelligence(AI).Method:A total of 493 cases of chest CT,including 248 CTs of COVID-19 from Wuhan No.1 Hospital and 245 CTs of CAP from the First Hospital of Jilin University,were retrospectively collected from January 2019 to April 2020.They were divided into two datasets:a training/validation set and a testing set,with an 80/20%split,the ratio of COVID-19 to CAP was 1:1 in each group.A deep learning model was trained and validated on 394 CT using 5-fold cross-validation.The model was further tested on an independent set of 99 CT.Furthermore,eight clinicians,who were blinded to the gold standard and the deep learning model's results,independently evaluated the test data.The accuracy,sensitivity,specificity,and area under the curve(AUC)were used to assess the diagnostic performance of the deep learning model and the eight clinicians.Results:(1)In the training/validation set,the deep learning model's average diagnostic performance was that the accuracy,sensitivity,specificity,and AUC were 94.7%,95.9%,93.1%,and 0.95,respectively.(2)In the independent test set,the deep learning model's diagnostic performance was that the accuracy,sensitivity,specificity,and AUC were 93.9%,96.0%,91.8%,and 0.94,respectively.(3)In the independent test set,the eight clinicians' average diagnostic performance was that the accuracy,sensitivity,specificity,and AUC were 64.3%?47.3%?81.6%? 0.64.The deep learning model's diagnostic performance was superior to that of the eight clinicians(P<0.05).Conclusion:The deep learning model could more accurately differentiate COVID-19 from CAP than the eight clinicians.Thus,it might assist clinicians in making a faster and more accurate diagnosis of COVID-19,thereby allowing for timely isolation of infected patients and slowing the spread of this disease.Innovation points:(1)Using advanced deep learning technology to extract complex abstract features of CT data,and establishing a model to differentiate COVID-19 from CAP automatically.(3)Comparing the deep learning algorithm and clinicians' diagnostic performance,and explore the auxiliary role of the deep learning algorithm in clinician's diagnosis.