The Application Of Convolutional Neural Network In Automatic Diagnosis Of Capsule Endoscopy Images

Background & Aims: Screening for gastric diseases in symptomatic outpatients with conventional esophagogastroduodenoscopy(C-EGD)is expensive and has poor compliance.We aimed to explore the efficiency and safety of magnetic-controlled capsule gastroscopy(MCCG)in symptomatic outpatients who refuse C-EGD.Methods: We performed a retrospective study of 76794 consecutive symptomatic outpatients from January 2014 to October 2019.A total of 2318 adults(F/M = 1064/1254)in the MCCG group who refused C-EGD were matched with adults in the C-EGD group using propensity-score matching(PSM).The detection rates of abnormalities were analyzed to explore the application of MCCG in symptomatic patients.Results: Our study demonstrated a prevalence of gastric ulcers(GUs)in patients with functional dyspepsia(FD)-like symptoms of 8.14%.The detection rate of esophagitis and Barrett’s esophagus was higher in patients with typical gastroesophageal reflux disease(GERD)symptoms than in patients in the other four groups(P <0.01).The detection rates of GUs in the five groups(abdominal pain,bloating,heartburn,follow-up and bleeding)were significantly different(P = 0.015).The total detection rate of gastric ulcers in symptomatic patients was 9.7%.A total of 7 advanced carcinomas were detected by MCCG and confirmed by endoscopic or surgical biopsy.The advanced gastric cancer detection rate was not significantly different between the MCCG group and the C-EGD matched group in terms of nonhematemesis GI bleeding(2 vs.2,P=1.00).In addition,the overall focal lesion detection rate in the MCCG group was superior to that in the CEGD matched group(224 vs.184,P=0.038).MCCG gained a clinically meaningful small bowel diagnostic yield of 54.8%(17/31)out of 31 cases of suspected small bowel bleeding.No patient reported capsule retention at the two-week follow-up.Conclusion: MCCG is well-tolerated,safe,technically feasible,and has considerable diagnostic yields.The overall gastric diagnostic yields of gastric focal lesions were comparable with C-EGD.MCCG offers a supplementary diagnosis in patients who had an undiagnostic C-EGD before,indicating that MCCG could play an important role in outpatients’ routine monitoring and follow-ups.MCCG shows its safety and efficiency in symptomatic outpatients’ appliance.Background & Aims: China is a country with a large number of digestive diseases.Convolutional neural networks(CNN)model are widely discussed by researchers due to its excellent image feature extraction performance.This project aims to use large sample prospectively tagged capsule endoscopy pictures to train a reliable and rapid intelligent CNN-based reading system.Hopefully,timely localization diagnosis and early intervention of emergency patients can improve patient prognosis.Methods: the images and baseline information of 1015 OGIB patients and 1373 nonOGIB patients visited the Ruijin Hospital affiliated to Shanghai Jiaotong University from August 2007 to July 2019 were used for model training,and all patients who planned to undergo capsule endoscopy for suspicious small bowel hemorrhage from February 2020 to July 2020 at Ruijin Hospital affiliated to Shanghai Jiaotong University were prospectively included as a second verification group.We trained a deep convolutional neural network(CNN)system based on Res Net50 structure.The sensitivity,specificity,and accuracy of the trained CNN model were calculated.The verification is divided into two stages.In the first stage,every single static image is used to initially analyze the diagnostic performance of the CNN mode.In the second stage,we prospectively enroll patients that meet the enrollment conditions,and use their complete video to verify trained CNN-based model.Results: The CNN-based system achieved an accuracy over 95% at different PS values in the task of distinguishing bleeding small bowel endoscope images from non-lesion(including mucus,air bubbles,bile,etc.)images.When the PS value sets 0.6,the sensitivity of the model is 98.46%,the specificity is 95.83%,and the overall accuracy is 96.04%.Only one positive picture is wrongly judged as negative.Conclusion: The CNN-based system has an advantage over traditional suspected blood indicator(SBI)in the differentiation of active small bowel bleeding images and nonlesions(including mucus,air bubbles,bile,etc.).The results show that the CNN model reaches very ideal sensitivity under different cutoff values,which proves that this model has strong clinical value and can be a screening tool for small intestinal bleeding.