| BackgroundEsophageal cancer is the common malignant tumor worldwide.Esophageal squamous cell carcinoma(ESCC),the predominant sub-type,is responsible for the greatest number of esophageal cancers in China.The prognosis of patients with esophageal squamous cell carcinomas is significantly correlated with tumor stage.The5 year survival rate of patients with advanced esophageal squamous cell carcinoma is less than 20%,while in early stage is above 90%.The early detection and treatment of ESCC can significantly improve patients' prognosis.However,the clinical manifestations of early-stage ESCC are not specific and the endoscopists lack of experience in detecting early esophageal lesions.In our country,the detection rate of the early-stage ESCC is less than 10%,which is far lower than in other countries.Therefore,an accurate and practical method is important for diagnosing early-stage ESCC.The narrow band imaging(NBI)system is the current highly sensitive method for detecting superficial ESCC.Further,the combination of NBI and magnifying endoscopy(NBI-ME)allows the visualization of lesion-associated intrapapillary capillary loops(IPCLs)and enables differentiation of the neoplastic and non-neoplastic tissues in clinical.The study aims to explore the clinical value of NBI-ME for IPCLs classification of early-stage ESCC.We also develop a computer-assisted diagnosis model to evaluate the feasibility of automated classification of IPCLs to improve the detection of the early-stage ESCC.Methods1.To explore the clinical value of NBI-ME for IPCLs classification of early-stage ESCC.We recruited patients who underwent NBI-ME examination for the evaluation of a suspicious esophageal condition at the Department of Gastroenterology,First Affiliated Hospital of Anhui Medical University in An Hui Provence,China,from December2013 to April 2018.The study aimed to explore the character of esophageal lesions,the relationship between classification of IPCLs and the histological findings,the relationship between endoscopic diagnosis and histological findings.2.Computer-assisted diagnosis based on deep learning for automatic detection of early-stage ESCC using narrow-band imaging magnifying endoscopy.1).Screening suitable NBI-ME images of early esophageal squamous cell carcinoma from the collected data and labeling typical areas.A CAD model with NBI-ME images was developed to evaluate the suspicious esophageal lesions.Fully convolutional networks with double-label for semantic segmentation were designed.The study was based on FCN_Alex Net and FCN_VGG16.We calculated the diagnostic accuracy of the CAD model at the pixel-level and the lesion-level,respectively.After two experts annotated the NBI-ME images with predefined colors using a computer drawing tool,images were randomly given to the model for training and testing according to the 3-fold cross validation principle,achieving the diagnostic accuracy of the model at the pixel-level.2).Then,the same test sets were given to the trained model and nine endoscopists for classification of the annotated area.In order to evaluate the performances of the model and nine endoscopists,the diagnostic accuracy of the model was also calculated at the lesion-level.The total number of lesions and the correctly classified lesions were recorded,achieving the diagnostic accuracy at the lesion-level.Results1.To explore the clinical value of NBI-ME for IPCLs classification of early-stage ESCC.Eventually,total 292 patients were enrolled for analysis.Among the ESCC patients confirmed by ESD or surgery,the diagnoses included low-grade intraepithelial neoplasia(LGIN,n=25)and early esophageal cancer(n=267).Among the early esophageal cancer,124 patients were diagnosed as high-grade intraepithelial neoplasia(HGIN).The diagnostic accuracy of early esophageal cancer was 92.5%.According to the classification of IPCL,234 patients were correctly evaluated with infiltration depth.The overall diagnostic accuracy was 80.1%(234/292).In addition,the diagnostic accuracy of type A,B1 and B2 was 92.5%,80.8% and 88.4%,respectively.Further analysis on the diagnostic value of IPCLs from 2013 to 2018,we found that diagnostic accuracy is proportional to the increase in endoscopic physician reviewer experience.Between 2013 and 2014,the overall diagnostic accuracy of classification of IPCLs was 65.8%.Between 2015 and 2016,the overall diagnostic accuracy of classification of IPCLs was 78.1%.Between 2017 and 2018,the overall diagnostic accuracy of classification of IPCLs was 87.2%.2.Computer-assisted diagnosis based on deep learning for automatic detection of early-stage ESCC using narrow-band imaging magnifying endoscopy1).Totally 219 patients were enrolled for analysis.Of the 1350 full-size images selected for analysis,206 were classified as type A,945 as type B1,and 199 as type B2.Segmentation results of whole images were based on FCN_Alex Net.The comparison showed the FCN-double model was close to the ground truth.The pixel-level diagnosis was made on ROIs,and the FCN-double model outperformed other FCN-single model.The best FCN-double model was based on VGG16 and achieved a segmentation result up to 77.8%.According to the 3-fold cross-validation principle.The diagnostic performance showed that the double-labeling FCN model achieved an average accuracy of 93.0% at the pixel level.Among the three test groups,the per-class mean diagnostic accuracy was the highest for IPCL type B1.A combined classification rate of87.0% at the lesion-level was obtained.2)Obvious group-based diagnostic discrepancies were found between the nine endoscopists and the gold standard when evaluating the whole 1383 lesions in the set.The best result was achieved by expert with NBI-ME inspection ?5 years' experience(group S),with an average diagnostic accuracy of 92.0%.The average diagnostic accuracy of was 82.0% in ?3 years' experience(group M),while the average diagnostic accuracy of was lower at 73.3% in ? 1 year(group J)of NBI-ME physicians.The diagnostic performance of the model was similar to the endoscopists in group S,achieving the average diagnostic accuracy of 89.2%.The sensitivity and specificity of the group S,group M and group J were 90.5% and 84.1%,78.6% and71.9%,67.7% and 76.4%,respectively.The sensitivity and specificity of the model were 87.0% and 84.1%.Further analysis revealed that for type A lesions,the diagnostic performance of the model was superior to that of endoscopists in groups M and J(P<0.05).These two groups had a low sensitivity for type A lesions and were apt to confuse them with type B1.Similarly,in type B lesions,although we found that the majority of endoscopists performed well when classifying IPCL type B2,there were significant deviations from the model(P<0.05).The model was more accurate than endoscopists in groups M and J for both B-lesion types(P<0.05).Moreover,the model's diagnostic accuracy for type B2 was not significantly different from that of the S group(P>0.05).The diagnostic performance was proportional to endoscopic experience.Group S closely matched the gold standard,with kappa values of 0.745–0.812.However,the interobserver agreement between the two other groups and the gold standard was unsatisfactory(kappa value,0.310–0.527).The interobserver agreement between the model and the gold standard was substantial(kappa value,0.719).Conclusions1.The AB classification of IPCL by NBI—ME can improve the diagnostic accuracy of the early-stage ESCC.However,the classification of IPCLs is highly operator dependent and shows a significant variability among the operators.Moreover,the diagnostic accuracy relies on the operator's experience.2.We proposed an FCN with double-label for semantic segmentation within the framework self-transfer learning.The double-labeling FCN model achieved an average accuracy of 93.0% at the pixel level and 89.2% at the lesion level.The model's accuracy was comparable to group S and significantly higher than that of groups M and J.The double-labeling FCN automated IPCL recognition is feasible and could facilitate early detection of ESCC. |
PDF Full Text Request