Development And Validation Of New Screening Methods For Esophageal Cancer Based On A Novel Esophageal Cell Enrichment Device And Deep Learning

China has the world heaviest burden of esophageal cancer,with more than half of the global esophageal squamous cell carcinoma(ESCC)incidence and a stubbornly high mortality.The five-year survival rate for ESCC patients is less than 30%.Meanwhile,the incidence of adenocarcinoma of the esophagogastric junction(AEG)in China is gradually increasing,which further increases the severity and complexity of esophageal cancer prevention and control.Population-based screening can increase the detection rate of early ESCC,AEG,and their precancerous lesions,thus reducing the mortality and incidence.Upper gastrointestinal endoscopy is currently the only well-established screening method for ESCC and AEG.However,its invasiveness and resource-intensiveness make it not feasible and cost-effective to implement in mass population.The lack of easy-to-use and accurate screening methods has long been a bottleneck in esophageal cancer prevention and control.Esophageal balloon cytology was utilized for ESCC and AEG screening in high-risk geographical areas in China last century,but its sensitivity was only about 40% and patient acceptability is poor.In recent years,international studies have achieved good results in Barrett’s esophagus screening by using esophageal sponge cytology combined with biomarker tests.However,such method has been very inadequately studied for ESCC and AEG,and some preliminary results suggested that the sensitivity is still lower than clinical expectations.These suggest that further research is needed to improve the existing methods for esophageal cell collection and subsequent detection,in order to improve its performance in ESCC and AEG screening.In this study,we improved the existing methods of esophageal cell collection and detection,and innovatively applied deep learning to assist cytological diagnosis and feature extraction.We established a new method for esophageal and gastroesophageal junction cancer screening and validated it in large-scale screening population.In the first part of this study,a novel esophageal cell enrichment device was developed to improve the performance of esophageal cell sampling.In the second part,a deep learning algorithm was established for identification of abnormal esophageal cells,by which the cytological reading speed and accuracy were significantly improved.In the third part,we evaluated the feasibility and diagnostic performance of the novel esophageal cell enrichment device and artificial intelligence(AI)assisted cytology for community-based screening of esophageal cancer.In the fourth part,we expanded the study population in order to develop an automated prediction model independent of cytopathologists.We integrated risk factors and digitalized cytological features extracted by deep learning algorithms,and trained and validated a multimodal machine learning model for prediction of ESCC and AEG.The new method is for primary screening and risk stratification in high-risk areas.It has the potential to reduce unnecessary endoscopies in low-risk individuals and allows concentration of high-risk individuals for endoscopy,thus improving the effectiveness and efficiency of ESCC and AEG screening programs.Part Ⅰ Development and sampling performance test of a novel esophageal cell enrichment deviceObjective: To develop a novel esophageal cell enrichment device with high cell collection quantity,good safety,and subject tolerance,and to observe and identify the collected cell components.Methods: The device was optimized and tested in terms of capsule material,cell collection device shape,and material parameters.During the in vitro experiment phase,the dissolution time of gelatin and glutinous capsule were tested at different water temperatures.Polyurethane sponge was chosen as the material for cell collection,and the time for expansion in water after packed in the capsule was evaluated in four designed shapes,including hemispherical,columnar,dome,and dome-trilobal.The capsule material and cell collection device shape with the shortest time of dissolution and expansion were selected,and cell collection device with different material pore densities were customized.Healthy volunteers were recruited for cell collection using devices with different parameters,and participant’s acceptability and adverse events of the sampling process were recorded.Fifty liquid-based cytology slides were made for each sample.Slides were scanned with a digital pathology system after Forgan-Eosin staining,and cell counting was finished automatically.The material parameter with the highest numbers of collected cells was selected.Morphology of collected cells was observed in liquid-based cytology slides to identify their histologic origins,and cell blocks were also made for assistance.Results: Glutinous capsule took the shortest time to dissolve in a 55°C water bath(39.6± 6.0s),and the dome-trilobal shaped cell collection device(60.8 ± 4.2s)expanded significantly faster than the other three shapes(P<0.001).The maximum number of collected cells in healthy volunteers was reached at the device material pore density of 45 pores per inch,and the maximum cell count was 6,685,000 ± 294,000.No severe adverse events were documented during the sampling process,and 63.3% of subjects reported only mild discomfort.The majority of collected cells were squamous epithelial cells,with scattered inflammatory cells,gastric-type glandular cell clusters,and tonsillar mucosal pieces.Conclusion: A novel esophageal cell enrichment device was developed.The new device can effectively collect cells from the esophagus,gastroesophageal junction,and part of the oropharynx with large cell counts,fast process,and good safety and tolerance.Part Ⅱ Development of an esophageal malignant cell classification model based on deep learningObjective: To develop an classification model for abnormal cells in esophageal cytology samples based on deep learning algorithms,and to improve the efficiency and accuracy of cytological diagnosis.Methods: Cytology samples were collected from patients with ESCC,AEG,highgrade intraepithelial neoplasia,and healthy volunteers with the novel esophageal cell collection device,and 50 liquid-based cytology slides were made for each sample in the same way as part I.An expert workshop was held to develop the cytological diagnostic classification and criteria for intraepithelial lesions of esophagus and gastroesophageal junction.On this basis,the esophageal cytological images were annotated as the gold standard for model training,of which the results were confirmed by at least two cytopathologists.The 89×89 pixels cell image tiles were extracted and assigned to the training and testing set in terms of individual subjects,so that the percentage of tiles in the two sets were roughly 75% and 25%,respectively.The images in training set were input into Visual Geometry Group network(VGG-16)deep residual network(Res Net-50)model.The training was completed until the loss functions of the training and testing sets no longer decrease,and the accuracy of classification was evaluated.A random selection of 100 positive and 100 negative slides was used to test the speed and accuracy of the cytopathologists’ review with and without AI assistance in a blinded state.Results: The testing set included 51,100 cytology tiles labeled as positive and 74,900 negative tiles,and the classification accuracy of the Res Net-50 model was significantly higher than that of the VGG-16 model(p<0.001).The testing set consisted of 51,100 cell tiles labeled as positive and 74,900 tiles as negative.The accuracy of the model for cell image classification was 98.3%(95% confidence interval [CI],98.2-98.4%).The sensitivity was 98.0%(95% CI,97.9-98.1%),and the specificity was 98.5%(95% CI,98.4-98.6%).Compared to manual review alone,the cytopathologists took significantly less time per slide under AI assistance(0.73 min vs 3.46 min,P<0.001)and the sensitivity for abnormal cells was also significantly improved(100.0% vs 89.0%,P=0.001).Conclusion: The trained deep residual network can accurately classify the benign and malignant cytological images,and improve the speed and sensitivity of cytopathologist in slides reading.Part Ⅲ Feasibility and diagnostic accuracy of artificial intelligence-assisted cytology for community-based screening of esophageal cancerObjective: To evaluate the feasibility and accuracy of the novel esophageal cell enrichment device and AI-assisted cytology for community screening of esophageal cancer.Methods: Residents aged 40–85 years were recruited in a high-risk area of ESCC.Esophageal cells were collected using an approved novel capsule sponge,and cytology slides were scanned by a trained AI system before cytologists provided confirmation.Atypical squamous cell or more severe diagnosis was defined as positive cytology.AI-based abnormal cell counts were also reported.Upper endoscopy was performed subsequently with biopsy as needed.Diagnostic accuracy,adverse events,and acceptability of cytology testing were assessed.Esophageal high-grade lesions(carcinoma and high-grade intraepithelial neoplasia)were the primary target lesions.Results: In total,1,844 participants were enrolled,and 20(1.1%)high-grade lesions were confirmed by endoscopic biopsy.The AI-assisted cytology showed good diagnostic accuracy,with a sensitivity of 90.0%(95% CI,76.9%–100.0%),specificity of 93.7%(95%CI,92.6%–94.8%),and positive predictive value of 13.5%(95% CI,7.70%–19.3%)for detecting high-grade lesions.The area under the receiver operation characteristics curve was0.926(95% CI,0.850–1.000)and 0.949(95% CI,0.890–1.000)for AI-assisted cytology and AI-based abnormal cell count,respectively.The numbers of endoscopy could be reduced by92.5%(from 99.2 to 7.4 to detect 1 high-grade lesion)if only cytology-positive participants were referred to endoscopy.No serious adverse events were documented during the cell collection process,and 96.1% participants reported this process as acceptable.Conclusion: The AI-assisted cytology is feasible,safe,and acceptable for ESCC screening in community,with high accuracy for detecting esophageal high-grade lesions.Part Ⅳ Construction and validation of a multimodal machine learning prediction model for esophageal and gastroesophageal junction cancer based on cytological images and macroscopic risk factorsObjective: To develop and validate a multimodal machine learning model integrating cytological features and macroscopic risk factors for large-scale population-based ESCC and AEG screening,reducing the dependence on cytopathologists.Methods: For this multicohort prospective study,we enrolled participants aged 40–75years undergoing upper gastrointestinal endoscopy screening at 39 tertiary or secondary hospitals in China for model training and validation(validation set I),and included community-based screening participants for further validation(validation set II).All participants underwent questionnaire surveys,sponge cytology testing,and endoscopy in a sequential manner.We trained machine learning models to predict a composite outcome of high-grade lesions,defined as histology-confirmed high-grade intraepithelial neoplasia and carcinoma of the esophagus and esophagogastric junction.The predictive features included105 cytological features extracted by deep learning and 15 epidemiological features.Model performance was primarily measured with the AUROC and average precision.Results: Between Jan 1,2021,and June 30,2022,17 498 eligible participants were involved in model training and validation.In the validation set I,the Light GBM classifier reached the best performance and was selected as the final model.The AUROC of the final model was 0·960(95% CI 0·937 to 0·977)and the average precision was 0·482(0·470 to0·494).The model achieved similar performance to consensus of cytologists with AI assistance(AUROC 0·955 [95% CI 0·933 to 0·975];p=0·749;difference 0·005,95% CI,–0·011 to 0·020).If the model-defined moderate-risk and high-risk groups were referred for endoscopy,the sensitivity was 94·5%(95% CI,88·8 to 97·5),specificity was 91·9%(91·2to 92·5),and the predictive positive value was 18·4%(15·6 to 21·6),and 90·3% of endoscopies could be avoided.Further validation in community-based screening showed that the AUROC of the model was 0·964(95% CI 0·920 to 0·990),and 92·8% of endoscopies could be avoided after risk stratification.Conclusion: We developed a prediction tool with favorable performance for screening of ESCC and AEG.This approach could prevent the need for endoscopy screening in many low-risk individuals and ensure resource optimization by prioritizing high-risk individuals.