Font Size: a A A

Study On Influential Factors Of Artificial Intelligence In Diagnosis Of GGN Invasion And Detection Of Pulmonary Nodules

Posted on:2021-05-19Degree:MasterType:Thesis
Country:ChinaCandidate:W ZhouFull Text:PDF
GTID:2404330623974073Subject:Imaging and nuclear medicine
Abstract/Summary:PDF Full Text Request
Purpose:To explore the value of deep learning-based artificial intelligence-assisted diagnosis system on the quantitative analysis of CT of lung ground glass nodule(GGN)for its pathological invasiveness judgment,and the effect of different layer thickness and nodule characteristics on the detection of small pulmonary nodules of CT influences.Materials and Methods:54 cases of lung adenocarcinoma ground glass nodules confirmed by surgery and pathology in our hospital from January 2018 to August 2019 and 200 cases of CT scan of small lung nodules from January 2019 to August 2019 were imported into artificial intelligence software Workstation for analysis.Forty-four GGN features recorded after analysis of the former include 3D long diameter,average CT value,volume,maximum face area,surface area,compactness,sphericity,and entropy.54 GGNs were classified as non-invasive lesions according to pathological classification.Group(AAH,AIS)and invasive disease group(MIA,IAC),SPSS and MedCalc software were used to statistically analyze the differences between the two groups,and then the receiver operating characteristic curves were performed for statistically significant quantitative parameters.(ROC)analysis to evaluate its ability to diagnose invasiveness of GGN.At the same time,the optimal diagnostic threshold of the quantitative parameter was calculated according to the Youden's index(YI),and the area under the curve,sensitivity and specificity were obtained,P<0.05 Differences were considered statistically significant.The latter used different CT image layer thickness,nodule type,nodule size,and model to record AI and physician detection of nodules,and SPSS 17.0statistical software was used for data entry,collation,and statistical analysis.Chi-square test was used to compare the qualitative data between the AI ??group and the doctor group.The t-test was used to compare the test time between the two groups.The detection rate,sensitivity,false positive rate,and detection time were compared.Result:1.There were 54 GGN in this study,of which 34 were pGGN,20 were mGGN,22(40.74%)were in the non-invasive lesion group(AAH,AIS),and 32(59.26%)werein the invasive lesion group(MIA,IAC).28 upper lobe of right lung(7 in the apical segment,11 in the posterior segment,10 in the front segment),5 in the right lung(1 in the inner segment,4 in the outer segment),and 8 in the lower lobe of the right lung(2 in the dorsal segment),6 anterior basal segments),10 in the upper left lobe(6 in the posterior cuspid segment,4 in the anterior segment),and 3 in the left lower lobe(1 in the dorsal segment and 2 in the basal segment).AI gave diagnostic opinions on 54 GGN as 48 GGNs were high-risk,accounting for 88.89%,and 6 were low-risk(4 pGGN,2mGGN),accounting for 11.11%.2.The 3D long diameter,average CT value,volume,maximum area,and surface area judged that the non-invasive lesion group was significantly different from the invasive lesion group.The optimal threshold for 3D long diameter diagnosis of invasive lesions was 10.35 mm,and the sensitivity was 84.37%,specificity is 59.09%,(AUC=0.740,P<0.05;95% CI: 0.603-0.850);the threshold of average CT value is476.73 Hu,sensitivity is 78.12%,specificity is 72.73%,(AUC=0.751,P<0.05;95% CI:0.615-0.859);volume threshold is 641.7mm3,sensitivity is 65.62%,specificity is81.82%,(AUC=0.737,P<0.05;95% CI: 0.600-0.848);the threshold of the maximum surface area is 79.78mm2,the sensitivity is 68.75%,and the specificity is 77.27%,(AUC=0.736,P<0.05;95% CI: 0.598-0.846);the threshold of the surface area is 365.22mm2,sensitive The specificity is 65.62% and the specificity is 81.82%,(AUC=0.744,P<0.05;95% CI: 0.607-0.853).3.There were no significant differences in the compactness,sphericity,and entropy of GGN between non-invasive and invasive adenocarcinoma(P> 0.05).4.Results A total of 755 true nodules were detected in 200 cases of chest CT,including pGGN218,mGGN61,and SN476.The detection rate of AI for pGGN and mGGN in thick layer(5mm)is higher than that in thin layer(1.5mm).According to statistical analysis,t=1.511,P= 0.134,t= 2.282,P= 0.025;AI is thin.The detection rate of SN in layers is significantly higher than that in thick layers(t=-10.377,P<0.001).The detection sensitivity of AI to pGGN,mGGN,and SN in thin layers is significantly higher than that in thick layers.According to statistical analysis,t=-4.823,P<0.001,t=-4.048,P<0.001,t=-10.186,P<0.001;the false positive rate of thin layers for solid nodules is significantly higher than thick layers,and thick layers for pure ground glass The false positive rate of nodules and mixed ground glass nodules was significantlyhigher than that of thin layers,and the physician's false positives were all 0.5.A total of 491 and 627 small pulmonary nodules were detected by AI in 64-row CT models and 16-row CT models.The two were statistically analyzed,t=-0.428,P=0.427,P> 0.05,no difference Statistical significance;small lung nodules were detected in both layers on thick layers than thick layers(t=-7.925,P<0.001);false positive rate of thin layers was significantly higher than thick layers(t=-2.890,P=0.004),among which the false positive rate of 16-row models is slightly higher than that of 64-row models.
Keywords/Search Tags:Artificial intelligence, deep learning, lung adenocarcinoma, ground glass nodule, computer tomograph
PDF Full Text Request
Related items