The Application Of Deep Learning In The Evaluation Of Malignant Risk Of Pulmonary Nodules On Chest Computed Tomography

The evaluation of probability of malignancy of pulmonary nodules is very important for the clinical management of them.The traditional pulmonary nodule risk prediction models rely on qualitive features including semantic radiological features and clinical information,which need human interpretation and have certain inter-observer variability.Recently,many studies focused on the development of radiomics or deep learning model on pulmonary nodule risk evaluation,but the clinical application still had challenges.Purpose:1.To propose a practical strategy for the clinical application of deep learning algorithms,i.e.,Hierarchical-Ordered Network-ORiented Strategy(HONORS),and a new approach to pulmonary nodule risk evaluation in various clinical scenarios,i.e.,Filter-Guided Pyramid NETwork(FGP-NET).We further validated the performance of HONORS based on FGP-NET.2.To compare the performance of FGP-NET with traditional risk prediction models and a group of 126 radiologists to investigate the potential value of it.Materials and Methods:1.We developed and validated FGP-NET on the Jinling Hospital(JLH)dataset and National Lung Screening Trial(NLST)dataset,which included a collection of 2106 pulmonary nodules combining screened and clinically detected nodules at CT images.We further performed external test on an independent multi-center test set(n=341)from three clinical hospitals.The area under the curves(AUCs)were used to assess the performance of FGP-NET.On top of FGP-NET,we validated HONORS which was composed of two solutions.First,in the Human Free Solution,we used the high sensitivity operating point for screened nodules,but the high specificity operating point for clinically detected nodules.Second,in the Human-Machine Coupling Solution,we used the Youden point.2.We compared the performance of FGP-NET with traditional risk prediction models including Mayo model and Brock model in the JLH test set.Besides,a comparison study with a group of 126 skilled radiologists in the JLH test set was also conducted.The AUCs of the Mayo model and Brock model were calculated.The sensitivity and specificity of each radiologist was calculated and the average performance of 126 radiologists was also calculated.We further voted the decisions of the 126 radiologists for each nodule to simulate the majority opinion of radiologists.The inter-radiologist agreements among all 126 radiologists,radiologists' maj ority opinion and FGP-NET were evaluated using Cohen's ? coefficient(?w).Results:1.FGP-NET achieved AUCs of 0.969(95%CI:0.943-0.986)and 0.847(95%CI:0.804-0.883)for internal and external test.Specifically,the AUCs for the subsets of the external test set ranged from 0.890 to 0.942.HONORS-guided FGP-NET identified benign nodules with high sensitivity(sensitivity,95.5%;specificity,72.5%)in the screened nodules,and identified malignant nodules with high specificity(sensitivity,31.0%;specificity,97.5%)in the clinically detected nodules.These nodules could be reliably diagnosed without any intervention from radiologists,via the Human Free Solution.The remaining ambiguous nodules were also diagnosed with high performance,which however required manual confirmation by radiologists,via the Human-Machine Coupling Solution.2.In the JLH test set,FGP-NET achieved an AUC of(0.927;95%CI,0.857-0.969),which was significantly higher than those of the Mayo clinic model(0.684;95%CI,0.583-0.773;p<0.001)and Brock model(0.727;95%CI,0.629-0.811;p<0.001).In the comparison study,we evaluated the performance of the 126 radiologists by their average sensitivity(72.2 ± 15.1%)and specificity(71,7±15.5%).Accordingly,FGP-NET achieved a higher sensitivity(93.3%)but a relatively inferior specificity(64.0%).Inter-observer agreement among 126 radiologists was fair(?w=0.315 ±0.139).We found that the inter-observer agreement between the radiologists' majority opinion and FGP-NET(?w = 0.538)was moderate,which outperformed 78 radiologists.Conclusions:FGP-NET performed comparably to skilled radiologists in terms of evaluating the risk of pulmonary nodules.HONORS,due to its high performance,might reliably contribute a second opinion,aiding in optimizing the clinical workflow and reducing medical errors.