| Background:Colorectal cancer(CRC)was one of the most common causes of cancer-related deaths worldwide.It’s widely accepted that the prognosis of CRC are closely related to the stage of the diseases.Colonoscopy provides a cost-effective tool for detecting and removing colorectal adenomas,thus decreasing the incidence and mortality of CRC.Screening colonoscopy has been widely adopted for the early detection CRC,especially in high-risk populations.However,false-negative rates of colonoscopy decrease its effectiveness to prevent CRC.Colonoscopy was a highly operator-dependent procedure which led to substantial variation between endoscopists in detecting colorectal adenomas.Standardized approaches have been issued in several guidelines and expert consensus for optimizing colonoscopy examination.Quality control based on the standardized approaches can decrease the variations in the performance of endoscopists and improve the effectiveness of colonoscopy to detect colorectal adenomas.Unfortunately,routine quality control for each colonoscopy procedure is difficult to carry out because the conventional way of manual quality control is time consuming and troublesome.In absence of routine quality measurement for colonoscopy,a certain miss rate of colorectal adenomas is unavoidable,leading gastroenterology specialty societies to call for a practical and effective method for quality control in routine colonoscopy.An ideal quality control method is required for real-time supervision and assistance for endoscopists in detecting colorectal polyps in a more practical way rather than manual quality control.With the development of artificial intelligence(AI),increasing attention has been focused on the application of deep convolutional neural networks(DCNNs)in computer-aided systems for assisting detection and diagnosis of various gastrointestinal abnormalities.Previous studies showed that DCNNs trained with specific endoscopic images were capable of specialist-level gastrointestinal lesions recognition.Therefore,it seems that the image recognition system using DCNNs has a promising future in supporting colonoscopy quality control.However,the real clinical significance of AI in colonoscopy quality control needs to be fully investigated before widespread application.Objectives:1.To develop and test DCNNs-based models for quality control of colonoscopy2.To prospectively assess whether the automatic quality control system(AQCS)could improve quality of colonoscopy in real clinical practiceMethods:Part Ⅰ:Development and testing of deep neural network-based models for quality control of colonoscopyWe developed DCNN models for automatically timing the withdrawal phase,supervising withdrawal stability,evaluating bowel preparation,and detecting colorectal polyps.Models B and E identify the beginning and end of the withdrawal phase,respectively.Models BP and PD evaluate bowel preparation and polyp detection,respectively.Model S supervises withdrawal stability.Firstly,26758 stored colonoscopic images were retrospectively collected from patients who underwent routine colonoscopy at Qilu Hospital of Shandong University.According to purposes of different quality-control models,each endoscopic image was reviewed and labeled by at least two experienced endoscopists.After the process of image preparation,the collected endoscopic images were divided into training datasets,validation datasets and testing datasets.Training datasets were used for training the models and validation datasets were used for optimizing the parameters of models during the training phase.After training,we assessed the performance of models B,E,BP,PD and S via testing datasets,and further tested model S on 30 real colonoscopy videos of withdrawal phase.Part Ⅱ:Clinical assessment of automatic quality control system in colonoscopyTo investigate the impact of DCNN models in colonoscopy quality control,a randomized controlled study was conducted in Qilu Hospital of Shandong University from October 2018 to May 2019.Firstly,we integrated five trained colonoscopy quality-control models into AQCS.Next,consecutive patients were prospectively randomized to undergo routine colonoscopies with or without the assistance of AQCS.All patients and data analysis were blinded to study group assignment,while the endoscopists were not blinded to randomization status.The primary outcome of the study was the adenoma detection rate(ADR)in the AQCS and control groups.The secondary outcomes of the study included polyp detection rate(PDR),mean number of adenomas and polyps detected per colonoscopy,mean withdrawal time,and adequate bowel preparation rate.Results:Part Ⅰ:Development and testing of deep neural network-based models for quality control of colonoscopy4980 colonoscopy images were collected for testing models B,E,PD,BP and S.The accuracy,sensitivity,and specificity of model B were 96.94%,97.06%and 96.65%.The accuracy,sensitivity,and specificity of model E were 98.20%,97.43%and 99.03%.The accuracy,sensitivity,and specificity of model PD were 95.47%,96.04%and 94.46%.The accuracy,sensitivity,and specificity of model S for recognizing blurry frames were 94.50%,91.20%and 97.80%.The accuracy of model BP for identifying different Boston Bowel Preparation Scores ranged from 97.00%to 98.30%.The area under the receiver operating characteristic(ROC)curves of models B,E and PD were 1.00,1.00 and 0.99.The area under the ROC curves of model S for recognizing blurry frames were 0.99.The area under the ROC curves of model BP ranged from 0.98 to 1.00.The testing result of 30 real colonoscopy videos showed accuracy of model S for identifying withdrawal stability was 93.33%.Part Ⅱ:Clinical assessment of automatic quality control system in colonoscopyA total of 659 patients were enrolled and randomized to two groups.After exclusion,308 and 315 patients were analyzed in AQCS and control group,respectively.ADR of AQCS group was significantly higher than that of control group(28.90%vs.16.51%,P<0.001),and AQCS group detected more adenomas per procedure than that of control group(0.367 vs.0.178,P<0.001).Meanwhile,PDR of AQCS group was significantly higher than that of control group(38.31%vs.25.40%,P=0.001),and AQCS group detected more polyps per procedure than control group(0.575 vs.0.305,P<0.001).In addition,the mean withdrawal time was significantly longer in AQCS group compared with control group(7.03±1.01 minutes vs.5.68 ±1.26 mintes,P<0.001).Adequate bowel preparation rate was significantly higher in AQCS group compared with control group(87.34%vs.80.63%,P=0.023).Conclusions:1.The DCNN models achieved high accuracy,sensitivity,and specificity for quality control of withdrawal time,withdrawal stability,bowel cleansing,and colorectal polyps detection.2.DCNNs-based AQCS significantly improved quality parameters of colonoscopy in real clinical practice,including withdrawal time,bowel preparation quality,and detection of colorectal polyps and adenomas.Significance:This study validated the feasibility of DCNN models in quality control of colonoscopy,and demonstrated that the use of DCNNs-based AQCS could increase the detection of colorectal adenomas in real clinical practice.Meanwhile,this procedure-incorporated AQCS could assist endoscopists in adhering to established quality indicators and improve performance of endoscopists in time.More importantly,this study put forward a computer-aided endoscopic quality-control strategy in an effective and practical manner which could decrease the burden and expense of hospitals in the future. |
PDF Full Text Request