Application Of Needle-based Confocal Laser Endomicroscopy And Deep Learning Model In The Diagnosis Of Gastric Subepithelial Lesions

Background and aimsGastric subepithelial lesions(SELs)are common in clinical practice.Due to lack of specific clinical manifestations and laboratory tests,gastric SELs are usually encountered incidentally during routine upper endoscopy.Gastric SELs encompass a diverse group of conditions,including non-neoplastic lesions and neoplastic lesions ranging from benign neoplasms,potentially malignant neoplasms and malignant neoplasms.Gastric SELs are mostly hemispherical or spherical protuberant entities with normal overlying mucosa.Thus,white light endoscopy alone is insufficient for the differential diagnosis of various gastric SELs.Currently,endoscopic ultrasound(EUS)is the most common modality for characterization of SELs.EUS is helpful to estimate the location,size,layer of origin,shape and echogenicity of lesions.Due to lack of biopsy samples,EUS only achieved the diagnostic yields of 45.5%-66.3%.Recently,several diagnostic techniques,such as EUS-guided fine-needle aspiration(EUS-FNA)and EUS-guided trucut biopsy(EUS-TCB),have emerged as methods for samples of gastric SELs to theoretically overcome the disadvantages of EUS.However,current studies demonstrated that EUS-FNA had variable diagnostic yields of 38%-82%because of inadequate tissue acquisition,whereas EUS-TCB was limited by technical constraints such as angulation caused by the stiffness of the device.Needle-based confocal laser endomicroscopy(nCLE)is a novel endomicroscopic technique that can be inserted through a 19-gauge needle into lesions and enables in vivo real-time microscopic imaging of tissue during EUS,called "optical biopsy".Recent studies have revealed the diagnostic value of EUS-guided nCLE(EUS-nCLE)for pancreatic cystic lesions,solid pancreatic masses and malignancy lymph nodes.However,no investigation has yet applied nCLE in the diagnosis of gastric SELs.During the examinations,CLE enables to achieve black-and-white images similar to histopathological images.Thus,the endomicroscopists need to have enough pathological knowledge to identify CLE images.Moreover,when the endomicroscopists performing CLE procedures for long time,the missed and delayed diagnosis would occur due to visual fatigue.Thus,a new method is needed,which can enable immediate and accurate diagnosis of gastric SELs during nCLE examination.Recently,with the development of computer science,artificial intelligence based on deep learning has shown great potential application in the area of image recognition.Currently,deep learning models have been applied in identifying the medical images,such as histopathological images,radiological images and endoscopic images.Therefore,the aims of this study were as follows:1)to explore the features of nCLE imaging in intra-abdominal organs and tissues and to preliminarily assess the feasibility of nCLE for clinical application;2)to establish the nCLE criteria for gastric SELs,to prospectively assess the diagnostic efficacy of real-time nCLE for gastric SELs,and to systematically investigate the feasibility,safety and reliability of nCLE;3)to construct the deep learning model for diagnosis of gastric SELs on nCLE images,and to evaluate the diagnostic value of deep learning model for gastric SELs.MethodsPart Ⅰ:In vivo imaging of needle-based confocal laser endomicroscopy in rabbit modelsThe healthy experimental rabbits(Japanese big-ear rabbits)were given intravenous 3%pentobarbital sodium(1 ml/kg)for general anesthesia via ear vein,followed by cutting the abdomen and exposing intra-abdominal organs and tissues.The nCLE miniprobe was inserted through the 19-gauge puncture model into various intra-abdominal organs and tissues(omentum majus,liver,pancreas,and psoas major).The nCLE images were acquired and real-time sequences of respective location were recorded.Finally,nCLE videos and corresponding pathological findings were evaluated and identified by three experienced endomicroscopists and one gastrointestinal pathologist in an unblinded manner.Part Ⅱ:Clinical application of needle-based confocal laser endomicroscopy in the diagnosis of gastric subepithelial lesionsThis part was divided into 2 stages,including Stage Ⅰ(pilot study)and Stage Ⅱ(validation study).1)Stage I:pilot studyFrom May 2016 to November 2016,we prospectively recruited patients scheduled to undergo EUS for gastric SELs>2 cm at Qilu Hospital,Shandong University,according to inclusion criteria and exclusion criteria.EUS examinations were performed by an experienced endoscopist to assess the morphologic features of gastric SELs.After the EUS examinations,the preloaded needle was punctured into the lesion under EUS guidance and the tip of the miniprobe was adjusted to optimize nCLE imaging.nCLE videos were recorded for further post-procedure analysis.The gold standard diagnosis was based on surgical histopathology from patients who underwent surgical operation or endoscopic histopathology from patients who were performed by endoscopic resection.nCLE videos and corresponding pathological findings were openly evaluated by three experienced endomicroscopists and one gastrointestinal pathologist.The nCLE criteria for gastric SELs were established.2)Stage II:validation studyFrom December 2016 to April 2018,consecutive patients with gastric SELs>1 cm were prospectively enrolled at Qilu Hospital,Shandong University,according to inclusion criteria and exclusion criteria.EUS examinations were performed by an experienced endoscopist to assess the morphologic features of gastric SELs.Based on these characteristics on EUS,a presumptive diagnosis was made by the endoscopist.During nCLE imaging,in vivo real-time nCLE diagnosis was made by the endoscopist according to the established nCLE criteria for gastric SELs in pilot study.Endomicroscopic videos acquired were stored in a specific folder.After the procedure,any procedure-relative adverse events were evaluated.The final diagnosis was based on histological analysis of surgical or endoscopic specimens.A one-month wash-out period after EUS-nCLE procedures,all nCLE videos were reviewed for off-line diagnosis by the same endoscopist who had made the real-time diagnosis.To assess the image quality and the interobserver agreements,all nCLE videos were randomized and reviewed by three experienced endomicroscopists in a blind manner.The image quality of nCLE videos was assessed by three endomicroscopists using a five-point scale(1 = poor,2 = fair,3 = moderate,4 = good,and 5 = very good).Part Ⅲ:Application of deep learning model in the diagnosis of gastric subepithelial lesions on needle-based confocal laser endomicroscopy imagesFrom March 2016 to June 2018,nCLE images of gastric SELs were collected from the database at the digestive endoscopic center of Qilu Hospital,Shandong University.All nCLE images were divided into training dataset,validation dataset and testing dataset.Three experienced endoscopists reviewed nCLE images with histopathological information and divided nCLE images into three categories including spindle cell tumor,ectopic pancreas and carcinoma according to nCLE criteria for gastric SELs.Manual annotation was performed and all annotations followed the same protocol.The algorithm engineer further processed the labelled images by clipping the images,shrinking the images,normalization and whitening in order better to extract the features of nCLE images The deep learning model was constructed by using Inception ResNet V2 convolutional neural network architecture.The training dataset and validation dataset were used to train the deep learning model.After being trained,the model was evaluated by the testing dataset.Meanwhile,the same testing dataset was reviewed by two experienced endomicroscopists in a blind manner and the diagnostic time required was recorded.ResultsPart Ⅰ:In vivo imaging of needle-based confocal laser endomicroscopy in rabbit modelsA total of five experimental rabbits(Japanese big-ear rabbits;3 males/2 females)were enrolled to undergo nCLE imaging.nCLE was successfully performed in intra-abdominal organs and tissues of all rabbits.The microscopic structures of cells,glands and microvessels in the omentum majus,liver,pancreas and psoas major were visualized clearly by nCLE,respectively.Characteristics of various intra-abdominal organs and tissues were displayed on nCLE images,which was correlated well with histopathological findings,as follows:1)omentum majus:regular,transparent,round cellular structures with white vessels;overlapping cellular structures with 3-demension;2)liver:regular,radial,black cord-like structures with white cord-like vessels;3)pancreas:regular,dark,lobular structures with regular mesh-like vessels;4)a,psoas major on longitudinal section:regular,black,long ribbon-like structures with white vessels;b,psoas major on transverse section:regular,round structures with black outlines.Part Ⅱ:Clinical application of needle-based confocal laser endomicroscopy in the diagnosis of gastric subepithelial lesions1)Stage Ⅰ:pilot studyA total of 33 patients(14 males/19 females;mean age 55.6 ± 11.2 years)were enrolled to undergo EUS-nCLE for 33 gastric SELs(mean size 3.9±3.1 cm).All 33 patients successfully underwent EUS-nCLE procedures.Of these 33 gastric SELs,20 cases were completely resected by endoscopic resection,whereas 13 cases were surgically resected.The definitive histologic results were gastrointestinal stromal tumor(GIST,n = 14),ectopic pancreas(n = 8),leiomyoma(n = 6)and carcinoma(n =5).Additionally,among 14 cases of GIST,GIST malignancy potential was assessed as follows:very low or low risk(n = 9)and intermediate or high risk(n = 5).Of 5 patients with carcinoma,4 cases were suspicious gastric linitis plastica with negative endoscopic biopsies(final histological results showing poorly differentiated adenocarcinoma),and 1 case was gastric SELs with metastatic carcinoma.The nCLE criteria for gastric SELs were established as follows:1)a,GIST(very low or low risk):dense,light gray,and fascicular architecture without glandular structures;increased vessels with mild fluorescein leakage;b,GIST(intermediate or high risk):heterogeneous fascicular architecture;dilated and distorted vessels with heterogeneous fluorescein leakage;2)ectopic pancreas:regular,dark,lobular structures with "coffee beans" appearance;regular mesh-like vessels;3)leiomyoma:homogeneous,loose,dark gray,and interlaced or fascicular architecture without glandular structures;increased vessels without fluorescein leakage;4)carcinoma:atypical glands/irregular,dark cell aggregates;dilated and distorted vessels with heterogeneous fluorescein leakage.2)Stage Ⅱ:validation studyEUS-nCLE was performed for 61 patients(25 males/36 females;mean age 54.1 ±11.0 years)with 61 gastric SELs(mean size 3.0± 2.6 cm).In all,one patient didn’t complete EUS-nCLE procedures successfully due to puncture failure.Finally,60 patients were included in the diagnostic analysis.Based on the endoscopic specimens(n = 35)or surgical specimens(n = 26),the final pathological diagnoses were GIST(n= 19),ectopic pancreas(n = 17),leiomyoma(n = 13)and carcinoma(n = 12).With regard to GIST malignancy potential,13 cases of GIST were classified as very low or low risk and 6 cases were intermediate or high risk.Of 12 patients with carcinoma,8 cases were suspicious gastric linitis plastica with negative endoscopic biopsies(final histological results showing poorly differentiated adenocarcinoma),2 cases were gastric SELs with poorly differentiated adenocarcinoma,and 2 cases were gastric SELs with metastatic carcinoma.According to the nCLE criteria for gastric SELs,GIST could be diagnosed by real-time nCLE with 73.7%sensitivity,95.1%specificity,and 88.3%accuracy;ectopic pancreas with 100.0%sensitivity,97.7%specificity,and 98.3%accuracy;leiomyoma with 91.7%sensitivity,93.8%specificity,and 93.3%accuracy;and carcinoma with 83.3%sensitivity,95.8%specificity,and 93.3%accuracy.Similarly,the sensitivity,specificity and accuracy of off-line nCLE diagnosis for GIST were 84.2%,95.1%and 91.7%,respectively;for ectopic pancreas the values were 100.0%,100.0%and 100.0%,respectively;for leiomyoma the values were 83.3%,93.8%and 91.7%.respectively;and for carcinoma the values were 100.0%,97.9%and 98.3%,respectively.There were no significant differences between real-time nCLE and off-line nCLE for diagnosis of gastric SELs(P>0.05).Moreover,the diagnostic accuracy and overall accuracy of real-time nCLE for gastric SELs were significantly higher than that of EUS alone(GIST,88.3%vs.70.0%,P =0.01;ectopic pancreas,98.3%vs.81.7%,P<0.01;leiomyoma,93.3%vs.75.0,P =0.01;carcinoma,93.3%vs.80.0%,P = 0.03;overall accuracy,86.7%vs.51.7%,P<0.01).Based on a five-point scale(1 = poor and 5 = very good),the mean image quality score of nCLE videos was 3.6 ± 0.1.The kappa values of interobserver agreements for GIST,ectopic pancreas,leiomyoma and carcinoma were 0.64,0.95,0.68 and 0.77,respectively.Part III:Application of deep learning model in the diagnosis of gastric subepithelial lesions on needle-based confocal laser endomicroscopy imagesA total of 6628 nCLE images of gastric SELs were collected,including 4371 training dataset images,1457 validation dataset images and 800 testing dataset images.In addition,the testing dataset included 260 nCLE images of spindle cell tumors(GIST and leiomyoma),243 nCLE images of ectopic pancreas,and 297 nCLE images of carcinoma.The area under the curve of the deep learning model for diagnosis of spindle cell tumors,pancreas and carcinoma was 0.97,0.99 and 0.95,respectively.At a threshold value of 0.33,spindle cell tumors could be diagnosed by deep learning model with 93.1%sensitivity,96.7%specificity,and 95.5%accuracy;ectopic pancreas with 97.9%sensitivity,98.2%specificity,and 98.1%accuracy;and carcinoma with 92.3%sensitivity,96.4%specificity,and 94.9%accuracy.The required time of deep learning model for diagnosing the testing dataset was 34 seconds.In contrast,the sensitivity,specificity and accuracy of the experienced endomicroscopist 1 and endomicroscopist 2 for spindle cell tumors were 94.6%and 93.9%,95.4%and 94.1%,95.1%and 94.0%,respectively;for ectopic pancreas the values were 96.7%and 97.5%,98.6%and 98.2%.98.0%and 98.0%,respectively;and for carcinoma the values were 90.6%and 88.2%,96.6%and 97.0%,94.4%and 93.8%,respectively.The required time of the experienced endomicroscopist 1 and endomicroscopist 2 for diagnosing the testing dataset was 2147 seconds and 2528 seconds,respectively.Except that the specificity of deep learning model and endomicroscopist 2 for diagnosing spindle cell tumors was significant(96.7%vs.94.1%,P = 0.04),there were no significant differences between deep learning model and two experienced endomicroscopists for gastric SELs(P>0.05).Moreover,there was no significant difference in the overall accuracy between deep learning model and two experienced endomicroscopists(94.3%vs.93.8%and 92.9%,P>0.05).Notably,compared with the experienced endomicroscopists,the diagnostic time of deep learning model was just 1.3%-1.6%of the former(34 seconds vs.2147 seconds and 2528 seconds).Conclusions1.nCLE images are correlated well with histopathological findings.2.Under EUS guidance,nCLE is feasible to visualized clearly the microscopic structures of cells,glands and microvessels in gastric SELs.Meanwhile,GIST,ectopic pancreas,leiomyoma and carcinoma can be identified by their specific nCLE images.3.According to the newly established nCLE criteria for gastric SLEs,nCLE is a rapid,safe and useful method for predicting gastric SELs in vivo during EUS,with a high degree of accuracy.4.The "substantial" or even "almost perfect" interobserver agreements are achieved for the differentiation of different types of gastric SELs,showing a high reliability of this classification system.5.Artificial intelligence can be applied to classify the nCLE images.The trained deep learning model based on convolutional neural network can rapidly diagnose gastric SELs with high accuracy.The diagnostic efficacy of deep learning model for gastric SELs is not inferior to the experienced endomicroscopists.Thus,deep learning model can be used as an assistant technique for the endoscopists to diagnose gastric SELs during nCLE imaging.SignificanceThis study demonstrates,for the first time,that nCLE is feasible to visualized clearly the microscopic structures of cells,glands and microvessels in diagnosing gastric SELs.nCLE can provide in vivo real-time diagnostic imaging and is a safe and reliable technique for predicting gastric SELs with a high degree of accuracy.Thus,the endoscopists can make spot-decision regarding whether to perform endoscopic resection or otherwise.Meanwhile,the trained deep learning model based on deep convolutional neural network can achieve a high accuracy for diagnosis of gastric SELs within a short time.The diagnostic efficacy of deep learning model for gastric SELs is not inferior to the experienced endomicroscopists.As a promising diagnostic method,deep learning model can be applied for training the nCLE trainees and assisting the endoscopists in diagnosing gastric SELs in the future.