Diagnostic value of magnification optical enhancement by experts and ResNet for detecting diminutive colorectal polyps: A retrospective study

Diagnostic value of magnification optical enhancement by experts and ResNet for detecting diminutive colorectal polyps: A retrospective study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Diagnostic value of magnification optical enhancement by experts and ResNet for detecting diminutive colorectal polyps: A retrospective study Xiaohan Wan, Yan Zhang, Huawei Yang, Yanqing Li This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4394161/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 12 You are reading this latest preprint version Abstract Background Deep learning models, such as narrow band imaging, are useful for identifying colorectal polyp pathological types using electronic chromoendoscopy. However, such findings remain unreported for i-scan optical enhancement (OE). Thus, this study proposes a ResNet-based model to distinguish between colorectal polyp types. Methods Endoscopic videos of 58 patients with colorectal polyps were collected, and their pathological types were determined by pathological examination. An experienced endoscopist was asked to predict the polyp pathological types shown in the videos to evaluate the diagnostic accuracy of the endoscopists. The same videos were extracted as images to train the ResNet model and evaluate its effectiveness. Finally, 100 endoscopic images were randomly selected for human–machine comparisons to examine the consistency between machine learning and endoscopist diagnoses. Results An accuracy of 89.77% was obtained by experienced endoscopists in distinguishing colorectal polyp types based on the entire video; the accuracy of the ResNet model based on video-extracted images was approximately 95%. Therefore, the ability of the ResNet model to classify colorectal polyps matched that of an experienced endoscopist. Conclusions I-scan OE provides a powerful tool for experienced endoscopists to differentiate between colorectal polyp pathological types; the ResNet model can help endoscopists achieve high diagnostic accuracy and reduce unnecessary invasive examinations. i-scan OE ResNet colorectal polyps pathological types Figures Figure 1 Figure 2 Figure 3 Figure 4 Background Colorectal cancer (CRC) is a highly prevalent malignant tumor in clinical practice, with an estimated annual incidence of approximately 1.93 million cases worldwide [ 1 ]. China accounts for a significant proportion of these cases, with an estimated 555 550 new cases reported annually. Fortunately, colonoscopic screening effectively reduces the incidence of CRC [ 2 ]. However, white-light endoscopy has poor efficacy in differentiating various colonic polyp types [ 3 ]. Hence, an invasive examination should be performed if any abnormality is detected, which can severely increase patient suffering. Such disease types can reportedly be identified in real time according to optical images of electronic chromoendoscopy, such as narrow band imaging (NBI) and optical enhancement (OE), which avoid unnecessary histological examinations of small adenomas and resection of small polyps [ 4 , 5 ]. However, the diagnostic efficacy of inexperienced endoscopists often struggles to achieve thresholds of ‘resect-and-discard’ and ‘diagnose-and-leave’ strategies outlined by the Preservation and Incorporation of Valuable Endoscopic Innovations (PIVI) [ 6 – 8 ]. Fortunately, recent advances in artificial intelligence (AI) have recognised the potential of computer-assisted algorithms for identifying images, which have been widely used to detect diabetic retinopathy and skin cancer [ 9 – 11 ]. Computer-assisted algorithms can help endoscopists make accurate diagnoses by providing real-time feedback and guidance and can overcome the limitations of high interobserver variability in polyp identification, leading to better patient outcomes and reduced healthcare costs [ 12 ]. However, there is a lack of deep learning models for predicting colorectal polyp pathological types during i-scan OE endoscopy, leading to large differences in the prediction of polyp pathological types among different endoscopists [ 13 ]. Therefore, the PIVI strategy becomes difficult to implement, resulting in increased medical expenses. Herein, we aimed to evaluate the accuracy of i-scan OE with magnification endoscopy for real-time differentiation of colorectal polyp pathological types and assessed whether it can help experts meet ‘resect-and-discard’ and ‘detect-and-leave’ strategy requirements [ 8 ]. Furthermore, a computer-assisted ResNet algorithm was trained to automatically differentiate between colorectal polyp pathological types. The diagnostic results of ResNet and AlexNet were compared to verify the superiority of ResNet over the conventional convolutional neural network (CNN) in colonoscopic image-based disease diagnosis. Methods Patient recruitment To evaluate the accuracy of i-scan OE with magnification endoscopy for real-time differentiation of colorectal polyp pathological types, patients were recruited between August 2016 and February 2018 at Qilu Hospital, Shandong University. The inclusion criteria were patients who underwent colonoscopy under anaesthesia, provided written informed consent, and had at least one polyp measuring < 10 mm. Patients with inflammatory bowel disease, CRC, or hereditary polyposis and those with incomplete colonoscopy or inadequate bowel preparation were excluded. Real-time study design and endoscopy procedure All videos were captured using a colonoscope equipped with an i-scan OE (EPK-i7010, Pentax, Japan). A MagniView EC-3890FZI magnifying endoscope with a 4–100 mm focal range and a maximum optical amplification multiple of 136 was used. Endoscopic operations were performed by two endoscopists using a rubber cap (Pentax Distal Rubber Hood OE-A59, Pentax, Japan). After bowel cleaning with 2 L polyethylene glycol, the patients underwent colonoscopy under intravenous propofol anaesthesia. For polyp identification, the standard protocol involved initiating the washing process and simultaneously recording a video of the procedure. The endoscopist used magnified mode 1 to predict polyp pathology and simultaneously recorded the videos. Diagnosis of pathological types Our preexperimental findings suggested promising diagnostic efficacy of the Japanese NBI Expert Team (JNET) classification in distinguishing between adenomas and hyperplastic polyps when observed under magnification i-scan OE. Therefore, here, we adopted the JNET classification as the standard for differentiation [ 14 ]. Lesions classified as normal, hyperplastic, or sessile serrated polyps were categorised as type I, characterised by the absence of visible blood vessels on their surface. In cases where the vessels were visible, the calibres of the vessels in the polyps were similar to those in the surrounding normal mucosa. Low-grade intramucosal neoplasia, including intramucosal cancer with low-grade structural atypia, was classified as type IIA, exhibiting dark brown microvessels that are uniform and arranged in a relatively well-ordered reticular pattern. An experienced endoscopist determined the pathology of each video clip. After observing the polyps, a biopsy was performed to obtain tissue samples for pathological examination. In cases wherein polyps were found in adjacent segments with similar surface structures, only one polyp was biopsied for further pathological examination. Polyp pathological types were categorised in accordance with the Vienna Classification into nontumorous (including inflammatory and hyperplastic polyps) and tumorous (such as tubular adenomas, tubulopapillary adenomas, and tubular papillary adenomas) polyps [ 15 ]. Patients with hamartomas were excluded from this study. Statistical analysis Statistical information regarding patient age, sex, polyp location, polyp morphology, polyp size, and final pathological type was collected and analysed. The polyp location was determined on the basis of the anatomical location. The duodenum and jejunum were classified as parts of the ascending colon, the ileum and jejunum were classified as parts of the transverse colon, and the sigmoid colon and rectum were classified as parts of the descending colon. Polyp morphology was classified in accordance with the Paris classification: flat-elevated (Ip), elevated (Is) and flat-depressed (IIs) [ 14 ]. Polyp size was estimated by endoscopists using biopsy forceps. The distance between the polyp and colon wall was used to evaluate polyp size. To assess the diagnostic value of polyps measuring ≤ 5 mm under magnification i-scan OE, we compared the consistency between the optical diagnosis results and pathological results for these polyps. The corresponding negative predictive value (NPV) was determined to evaluate whether the criteria for PIVI were met. To assess the performance of the diagnostic process, several statistical measures were calculated, including sensitivity, specificity, positive predictive value (PPV), NPV, and accuracy. The statistical software packages NCSS 11 and R (version 3.3.2) were used to perform the calculations. Post hoc diagnosis with ResNet Video files of the i-scan OE were first cut into images because the ResNet constructed here only receives pictures as input. Each second of the videos contained 30 pins, five of which were selected at equal intervals. The images contained three parts, labelled regions A (tissue district), B (edge district) and C (supplementary information district) (Fig. 1(a)). Because only region A contained useful information for classification, the other two regions were removed before further operations. The after-tailored pictures have a pixel dimension of 932 × 702, which is sufficient to contain pathological and partially normal tissues (Fig. 1(b)). However, many pictures were repetitive because of the irregular and slow camera movement; therefore, a process of deduplication was needed. All pictures were resised to 10 × 10 pixels according to the regional means. The D-hash was obtained by comparing the red-channel values of the pixels with those of adjacent pixels. The result was 0 if the right one was larger than the left one; otherwise, it was 1. Adjacent pictures with the same D-hash sequence were deemed the same, and only one picture was reserved. Finally, the blurry images were artificially removed by two experienced endoscopists to further guarantee sample quality. Typical neoplastic polyp and hyperplastic polyp samples are illustrated in Fig. 1(c) and (d), respectively. Construction of the deep learning model An 18-layer ResNet was adopted as the classification model; its framework is shown in Fig. 2 . Extending the conventional CNN, a mass of residual blocks is contained in ResNet, whose input can be transmitted to its output through a branch consisting only of a one-dimensional convolution layer. ResNet partially overcomes the problem of a disappearing gradient in a deep CNN and possesses a stronger fitting and generalisation ability. Notably, three-channel RGB images were chosen as the input to the network to preserve the most information. Although the complexity of the network was greater than that of the grayscale image, there has been significant progress in diagnostic accuracy. The image dataset was randomly divided into training and test sets at a ratio of 8:2. Weighted sampling was used to mitigate the effects of sample imbalance. For ResNet, the key parameters included the size of the convolution kernel and the mapping size of each layer. Particle swarm optimisation was introduced to optimise these parameters. To simplify the calculations, the parameters of layers other than the first layer were set at a constant ratio; therefore, there were only three parameters to be optimised: the kernel size of the first layer, the kernel size of other layers, and the padding size. The optimisation goal was to maximise the average loss on the test set after training for 100 epochs; the training time was simultaneously used as an important item. The ultimate objective function is expressed as follows: where r is the reward, and a higher value indicates a more effective parameter. The AUC denotes the area under the receiver operating characteristic (ROC) curve of the diagnostic model after training; train is the training time in hours for 100 epochs, and b is the bias, which was set to 10 in our experiment. Results Patients and polyps Between March 2017 and August 2017, a dataset comprising 88 polyps from 58 patients (Table 1 ) was collected. Table 1 summarises the polyp characteristics. Table 1 Characteristics of the 88 polyps in the dataset Total (n = 88) Neoplastic (n = 51) Hyperplastic (n = 37) Size ≤ 5 mm 71 38 33 6–10 mm 17 13 4 Location Rectum 20 11 9 Sigmoid Colon 16 9 7 Descending Colon 18 10 8 Transverse Colon 17 10 7 Ascending Colon 12 6 6 Cecum 5 5 0 Morphology Is + Ip 66 44 22 IIa 22 7 15 Pathology 51 37 Tubular Adenoma 49 - Tubulovillous Adenoma 2 - Diagnostic performance of experienced endoscopists Regarding the diagnostic performance of the deep neural network and endoscopists in differentiating between neoplastic and hyperplastic colorectal polyps (Table 2 ), experienced endoscopists exhibited 89.77% accuracy, 96% sensitivity for adenoma identification, 81.58% specificity, 93.94% NPV, and 87.27% PPV during radiographic recording for the 88 polyps. Table 2 Diagnostic performance of the deep neural network and endoscopists in differentiating colorectal polyp types Sensitivity, n (%) Specificity, n (%) Accuracy, n (%) PPV, n (%) NPV, n (%) Realtime Diagnosis Experienced Expert 1 48/50 (96.00) 31/38 (81.58) 79/88 (89.77) 48/55 (87.27) 31/33 (93.94) Post hoc Diagnosis Expert 1 77/89 (86.52) 5/11 (45.45) 82/100 (82.00) 77/83 (92.77) 5/17 (29.41) Expert 2 75/89 (84.27) 7/11 (63.64) 82/100 (82.00) 75/79 (94.94) 7/21 (33.33) Experienced Expert 2 81/89 (91.01) 8/11 (72.73) 89/100 (89.00) 81/84 (96.43) 8/16 (50.00) ResNet 83/89 (93.26) 8/11 (72.73) 91/100 (91.00) 83/86 (96.51) 8/14 (57.14) Expert 1 with ResNet 81/89 (91.01) 7/11 (63.64) 88/100 (88.00) 81/85 (95.29) 7/15 (46.67) Expert 2 with ResNet 77/89 (86.52) 9/11 (81.82) 86/100 (86.00) 77/79 (97.47) 9/21 (42.86) Experts 1 and 2 were inexperienced endoscopists, whereas Experienced Experts 1 and 2 were experienced endoscopists. Image analysis by the trained ResNet model The entire training process, comprising 1000 epochs, took approximately 9.5 hours. By considering adenomatous polyps as positive examples and hyperplastic polyps as negative examples, curves for the accuracy, sensitivity, specificity and NPV of the testing data during training are depicted in Fig. 3(a), (b), (c), and (d), respectively. The accuracy reached a stable value of approximately 95%, with a corresponding sensitivity, specificity and NPV of 97.5%, 94% and 91%, respectively. Thus, the proposed method demonstrates favourable training performance, achieving a balance between sensitivity and specificity. Fig. 4 presents the ROC curves for distinguishing hyperplastic polyps from adenomas using the CNN model. The AUC was 0.98, which indicates that our model has good diagnostic performance in accurately identifying and differentiating colorectal polyps. Comparison between ResNet and endoscopists A randomly selected set of 100 images was prepared from a dataset outside the training and testing sets, consisting of both hyperplastic polyps and adenomas. This study evaluated the diagnostic capabilities of ResNet and experienced endoscopists. Two inexperienced endoscopists and one experienced endoscopist observed images of colorectal polyps and differentiated adenomas from hyperplastic polyps. In parallel, the trained ResNet model predicted polyp histological type. Five days later, the two experienced endoscopists differentiated polyps again with the help of the ResNet model. The diagnostic performance of ResNet was comparable to that of experienced endoscopists (Table 2 ). With AI assistance, inexperienced endoscopists demonstrated similar abilities in differentiating between the two histological polyp types. Comparison between ResNet and other networks To further confirm the advantages of our diagnostic neural network, we implemented an AlexNet model to perform the same task. As a traditional CNN model, AlexNet is a well-known deep learning model in the field of computer vision and is often used for disease diagnosis. The accuracy of the AlexNet model is shown in Additional File 1(a). The final accuracy achieved on the test set was approximately 91.5%, which required approximately 1000 epochs to reach. Additionally, as shown in Additional File 1(a), the accuracy of AlexNet on the training set reached 100% after 800 epochs, indicating that overfitting occurred in the model and that the effects were severely restrained. The sensitivity and specificity of the AlexNet model are presented in Additional File 1(b) and Additional File 1(c), respectively. In summary, the ResNet-based diagnostic model has advantages over conventional models in terms of its diagnostic effect and generalizability. Thus, our diagnostic neural network outperformed the conventional AlexNet model in terms of accuracy and convergence speed. Discussion In this study, involving 58 patients with colorectal polyps, an experienced endoscopist made real-time judgments regarding colorectal polyp pathological types. The i-scan OE optical diagnosis reached the PIVI threshold for detecting adenomas, achieving a high NPV of 93.94%. To address the low consistency of diagnoses made by different endoscopists, we developed a ResNet-based deep learning model to help doctors improve their diagnostic efficiency. The model showed excellent performance in distinguishing between hyperplastic and neoplastic polyps, with an accuracy of 95%, sensitivity of 97.5%, specificity of 94% and NPV of 91%, and it displayed a performance comparable to that of the diagnosis made by experienced endoscopists. Additionally, the model outperformed other deep learning models, such as AlexNet. Finally, we extracted another 100 images to test the model's ability. The overall accuracy rate was as high as 91%, which is comparable to that of an experienced endoscopist. To reduce the unnecessary need for pathological examination of colorectal polyps, two treatment strategies for PIVI have been proposed: ‘resect-and-discard’ and ‘detect-and-leave’. The diagnostic efficacy of PIVI can be met by experienced expert physicians in large medical centres; however, despite the help of electronic chromatic imaging, such as NBI, the results from primary care institutions and nonexpert endoscopists have been unsatisfactory [ 6 ]. The rapid development of machine learning has provided solutions for this problem. Recent studies have demonstrated that CNNs are effective in NBI-guided colorectal polyp identification. In 2017 and 2018, two studies were published, both of which showed significant progress in the use of deep learning models to identify colorectal polyp histopathology during NBI and achieved high sensitivity and accuracy [ 16 , 17 ]. With the advancement of deep learning models and new electronic endoscopic technologies, several recent studies have used various deep learning models to accurately differentiate among colorectal polyp pathological types [ 3 , 18 , 19 ]. In one study, researchers investigated the effectiveness of a CNN-based deep learning model in assisting endoscopists in differentiating colorectal polyp pathological types using the ELUXEO 7000 endoscopy platform [ 20 ]. The NPV of AI-assisted optical diagnosis for diminutive rectosigmoid polyps was 91.0%, meeting the required PIVI thresholds. Thus, AI can provide real-time guidance to endoscopists for the differential diagnosis of colorectal polyps based on endoscopic videos. The i-scan OE is a novel electronic chromoendoscopy system developed by the Pentax Corporation that incorporates a continuous light-wave enhancement filter in mode 1, which enhances the baseline projection and produces results similar to those of NBI. Although i-scan OE is widely used in many regions, few studies have been conducted on the potential of endoscopists and deep learning models to predict colorectal polyp pathological types using the i-scan OE system. Our model demonstrated a comparable differentiation ability for colorectal polyp pathological types in a retrospective analysis within approximately 1000 epochs of training, which is faster than that reported in previous research. Moreover, the sensitivity, specificity, accuracy and NPV of our model were similar to those of the aforementioned studies based on electron chromatic images. The rapid development of AI has led to the emergence of new deep learning algorithms suitable for image analysis, providing more choices. Herein, a new deep learning algorithm, ResNet, which has never been applied to colorectal polyp differentiation, was adopted. The ResNet model outperformed the AlexNet model, a traditional CNN model, in terms of diagnostic efficacy. After training on a training set of 480 images, ResNet effectively distinguished neoplastic polyps from hyperplastic polyps in the test set. This suggests that in the field of intelligent image classification for endoscopic images, we should follow the emergence of new deep learning models and continuously optimise existing image recognition systems. Our study is the first to evaluate the diagnostic efficacy of colorectal polyp histology using experts and a deep learning model. We are also among the first to apply the relatively less commonly used deep learning model ResNet to the field of colorectal polyp image identification by training an efficient ResNet model that can accurately identify polyp images. The results of the human–machine comparison show that the performance of two inexperienced endoscopists improved significantly upon the use of this model, demonstrating that our deep learning model can help improve the efficiency of endoscopic diagnosis after a short period of training. However, our study has several limitations. First, magnifying i-scan OE is not commonly used in colorectal endoscopy because transparent caps increase the difficulty of insertion. Only patients who underwent colonoscopy with a magnified endoscope were included here; therefore, polyps in these patients may not be representative of all patients with polyps. Moreover, we collected information from only 58 patients with colorectal polyps at Qilu Hospital. The collection of additional clinical data from multiple medical centers could validate the effectiveness of our model. Finally, the NPV and proportion of neoplastic polyps included in the study were related to some extent. Patient selection bias may also affect the NPV; therefore, further large-scale randomised controlled trials are needed to validate the diagnostic efficacy of magnifying i-scan OE for predicting the pathological type of colorectal polyps. Conclusions We developed a deep learning model for the diagnosis of colorectal polyp pathological type based on i-scan OE images with an NPV of 91%, which met the PIVI threshold. Further validation of the ability of this model in a multicentre prospective clinical trial is needed to enable its application to assist endoscopists, especially those who are untrained, in achieving the clinical practice requirements proposed by the PIVI within a short period, thereby changing the current situation wherein the nonselective resection of small colorectal polyps wastes a significant amount of medical resources and causes a significant economic burden for patients. List of abbreviations CRC, colorectal cancer; NBI, narrow band imaging; OE, optical enhancement; PIVI, preservation and incorporation of valuable endoscopic innovations; AI, artificial intelligence; CNN, convolutional neural network; JNET, Japanese NBI Expert Team; NPV, negative predictive value; PPV, positive predictive value; ROC, receiver operating characteristic; AUC, area under the curve Declarations Ethics approval and consent to participate: The current study was approved by the Medical Ethics Committee of Qilu Hospital of Shandong University (no. 2016043) and has completed the review process for registration on a clinical trials database, which can be accessed at https://clinicaltrials.gov/ with the unique identifier NCT02929641. All necessity for written informed consent from patients was waived by the Medical Ethics Committee of Qilu Hospital of Shandong University (no. 2016043 ), as long as patient data remained anonymous. The study adheres to the ethical principles outlined in the Declaration of Helsinki. Consent for publication: Not applicable. Availability of data and materials: Not applicable. Competing interests The authors declare no conflicts of interest for this article. Funding This study was supported by the National Natural Science Foundation of China (82270580 and 82070552), the Key Research and Development Program of Shandong Province (2021CXGC010506) and the Taishan Scholars Program of Shandong Province. Authors’ contributions: XW: conceptualisation and design; analysis and interpretation of the data; drafting of the article; and final approval of the article YZ: conceptualisation and design; analysis and interpretation of the data; and final approval of the article HY: analysis and interpretation of the data; critical revision of the article for intellectual content; and final approval of the article YL: conceptualisation and design; analysis and interpretation of the data; critical revision of the article for intellectual content; and final approval of the article Acknowledgements: Some of the graphic elements in the graphical abstract of this article were sourced from FigDraw (ID: USOISb27bb). References Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2021;71:209–49. Bretthauer M, Løberg M, Wieszczy P, Kalager M, Emilsson L, Garborg K, et al. Effect of colonoscopy screening on risks of colorectal cancer and related death. N Engl J Med. 2022;387:1547–56. Glissen Brown JR, Mansour NM, Wang P, Chuchuca MA, Minchenberg SB, Chandnani M, et al. Deep learning computer-aided polyp detection reduces adenoma miss rate: a United States multi-center randomized tandem colonoscopy study (CADeT-CS Trial). Clin Gastroenterol Hepatol. 2022;20:1499–e15074. Vișovan II, Tanțău M, Pascu O, Ciobanu L, Tanțău A. The role of narrow band imaging in colorectal polyp detection. Bosn J Basic Med Sci. 2017;17:152–8. Iacucci M, Trovato C, Daperno M, Akinola O, Greenwald D, Gross SA, et al. Development and validation of the SIMPLE endoscopic classification of diminutive and small colorectal polyps. Endoscopy. 2018;50:779–89. Rees CJ, Rajasekhar PT, Wilson A, Close H, Rutter MD, Saunders BP, et al. Narrow band imaging optical diagnosis of small colorectal polyps in routine clinical practice: the Detect Inspect Characterise Resect and Discard 2 (DISCARD 2) study. Gut. 2017;66:887–95. Klenske E, Zopf S, Neufert C, Nägel A, Siebler J, Gschossmann J, et al. I-scan optical enhancement for the in vivo prediction of diminutive colorectal polyp histology: results from a prospective three-phased multicentre trial. PLoS ONE. 2018;13:e0197520. Rex DK, Kahi C, O'Brien M, Levin TR, Pohl H, Rastogi A, et al. The American Society for Gastrointestinal Endoscopy PIVI (Preservation and Incorporation of Valuable Endoscopic Innovations) on real-time endoscopic assessment of the histology of diminutive colorectal polyps. Gastrointest Endosc. 2011;73:419–22. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542:115–8. Mookiah MR, Acharya UR, Chua CK, Lim CM, Ng EY, Laude A. Computer-aided diagnosis of diabetic retinopathy: a review. Comput Biol Med. 2013;43:2136–55. Takiddin A, Schneider J, Yang Y, Abd-Alrazaq A, Househ M. Artificial intelligence for skin cancer detection: scoping review. J Med Internet Res. 2021;23:e22934. Min JK, Kwak MS, Cha JM. Overview of deep learning in gastrointestinal endoscopy. Gut Liver. 2019;13:388–93. Iacucci M, Gasia M, Akinola O, et al. Mo1743 accuracy and inter-observer agreement for characterization of colonic polyps using the new optical enhancement -iSCAN Closefocus™ Colonoscope. Gastroenterology. 2016;150:S769. Shaukat A, Kaltenbach T, Dominitz JA, Robertson DJ, Anderson JC, Cruise M, et al. Endoscopic recognition and management strategies for malignant colorectal polyps: recommendations of the us multi-society task force on colorectal cancer. Gastroenterology. 2020;159:1916–e19342. Schlemper RJ, Riddell RH, Kato Y, Borchard F, Cooper HS, Dawsey SM, et al. The Vienna classification of gastrointestinal epithelial neoplasia. Gut. 2000;47:251–5. Byrne MF, Chapados N, Soudan F, Oertel C, Linares Pérez M, Kelly R, et al. Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during analysis of unaltered videos of standard colonoscopy using a deep learning model. Gut. 2019;68:94–100. Chen PJ, Lin MC, Lai MJ, Lin JC, Lu HH, Tseng VS. Accurate classification of diminutive colorectal polyps using computer-aided analysis. Gastroenterology. 2018;154:568–75. Gong EJ, Bang CS, Lee JJ, Seo SI, Yang YJ, Baik GH, et al. No-code platform-based deep-learning models for prediction of colorectal polyp histology from white-light endoscopy images: development and performance verification. J Pers Med. 2022;12:963. Pacal I, Karaboga D. A robust real-time deep learning based automatic polyp detection system. Comput Biol Med. 2021;134:104519. Rondonotti E, Hassan C, Tamanini G, Antonelli G, Andrisani G, Leonetti G, et al. Artificial intelligence-assisted optical diagnosis for the resect-and-discard strategy in clinical practice: the Artificial intelligence BLI Characterization (ABC) study. Endoscopy. 2023;55:14–22. Additional Declarations No competing interests reported. Supplementary Files FigureS1.zip GA.png Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 26 Nov, 2025 Reviews received at journal 25 Nov, 2025 Reviewers agreed at journal 10 Nov, 2025 Reviews received at journal 12 Mar, 2025 Reviewers agreed at journal 24 Feb, 2025 Reviewers agreed at journal 16 Feb, 2025 Reviewers agreed at journal 16 Dec, 2024 Reviewers invited by journal 15 Jul, 2024 Editor invited by journal 10 May, 2024 Submission checks completed at journal 10 May, 2024 Editor assigned by journal 10 May, 2024 First submitted to journal 09 May, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4394161","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":303512003,"identity":"9733b7e2-8f56-443d-a353-8332c713a683","order_by":0,"name":"Xiaohan Wan","email":"","orcid":"","institution":"Qilu Hospital of Shandong University","correspondingAuthor":false,"prefix":"","firstName":"Xiaohan","middleName":"","lastName":"Wan","suffix":""},{"id":303512004,"identity":"d8ff6d5f-435f-4f64-8d61-906197005666","order_by":1,"name":"Yan Zhang","email":"","orcid":"","institution":"Qilu Hospital of Shandong University","correspondingAuthor":false,"prefix":"","firstName":"Yan","middleName":"","lastName":"Zhang","suffix":""},{"id":303512005,"identity":"7bd6e6d4-319b-49a0-a45d-c147c6744fc2","order_by":2,"name":"Huawei Yang","email":"","orcid":"","institution":"Qilu Hospital of Shandong University","correspondingAuthor":false,"prefix":"","firstName":"Huawei","middleName":"","lastName":"Yang","suffix":""},{"id":303512006,"identity":"a6b08bab-c293-4b34-8d70-7de7a689d60e","order_by":3,"name":"Yanqing Li","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4ElEQVRIiWNgGAWjYBACPmYwJcHAwN4AIhkYGwhpYYNr4TlArBY4SyKBWC3sPIafC8os8uQjnz+8zcNgI7vhAPOzB/gdxmMsPeOcRLHh7Rxjax6GNOMNB9jMDQhoMZDmbZNI3Dg7h02ah+Fw4oYDPGwShGz5DdYy8/gzoJb/RGkxA9syX4LBDKjlADFa2Mqsec5JJG7gyTG2nGOQbDzzMJsZXi38/Ic33+Ypq0uc33784Y03FXayfcebn+HVwsDAYQCOHYMDIA4oqJjxqwcC9gdgLfINBFWOglEwCkbBSAUAzEM9jU5KTmcAAAAASUVORK5CYII=","orcid":"","institution":"Qilu Hospital of Shandong University","correspondingAuthor":true,"prefix":"","firstName":"Yanqing","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2024-05-09 09:35:31","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4394161/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4394161/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":56888352,"identity":"630a8bc3-34e7-42aa-89e2-800ac85f901b","added_by":"auto","created_at":"2024-05-21 19:01:17","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":999541,"visible":true,"origin":"","legend":"\u003cp\u003eImages of Japanese NBI Expert Team (JNET) type I and IIA polyps. These images were captured using i-scan optical enhancement (OE) technology.\u003c/p\u003e\n\u003cp\u003e(a) An untreated image was extracted from endoscopic videos and consisted of three parts: region A (tissue district), region B (edge district) and region C (supplementary information district).\u003c/p\u003e\n\u003cp\u003e(b) An image that was pretreated and used for model training.\u003c/p\u003e\n\u003cp\u003e(c) and (d) Magnified typical endoscopic images classified in accordance with the JNET classification for type I and IIA polyps captured by i-scan OE mode 1.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-4394161/v1/4440585d30d92002f35b1f02.png"},{"id":56888351,"identity":"f9f1dff6-28b0-4ace-9396-ce86cc792e05","added_by":"auto","created_at":"2024-05-21 19:01:17","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":191481,"visible":true,"origin":"","legend":"\u003cp\u003eFramework of the ResNet model\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-4394161/v1/49f117e96290f9db9dc2efa8.png"},{"id":56888353,"identity":"8d261664-2a87-40a8-bf21-e7d27a54c35e","added_by":"auto","created_at":"2024-05-21 19:01:17","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":256668,"visible":true,"origin":"","legend":"\u003cp\u003eDiagnostic performance of the ResNet model in differentiating between adenomas and hyperplastic polyps. Curves for accuracy (a), sensitivity (b), specificity (c) and NPV (d) of the testing data during training are depicted. NPV: negative predictive value.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-4394161/v1/ef1b51d725c3e83cfae36d03.png"},{"id":56888355,"identity":"4cd1e23d-5ce3-4c3d-af3a-ee5f7b440fa5","added_by":"auto","created_at":"2024-05-21 19:01:17","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":51415,"visible":true,"origin":"","legend":"\u003cp\u003eReceiver operating characteristic (ROC) curve for the model differentiation of adenomas versus hyperplastic polyps. AUC: area under the curve.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-4394161/v1/c7f5a12b8bdcd71a0b995975.png"},{"id":56889858,"identity":"d715c58d-49f0-43c4-888c-3344338d2b37","added_by":"auto","created_at":"2024-05-21 19:17:18","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1954027,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4394161/v1/9a6e0674-cd6b-44e1-8633-b68291141628.pdf"},{"id":56888356,"identity":"dfe2b860-f1b3-41d8-95ac-79d34b1b1a08","added_by":"auto","created_at":"2024-05-21 19:01:17","extension":"zip","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1166860,"visible":true,"origin":"","legend":"","description":"","filename":"FigureS1.zip","url":"https://assets-eu.researchsquare.com/files/rs-4394161/v1/f793a6a21a1bf84fab2c7a50.zip"},{"id":56889492,"identity":"34e68516-20c8-4751-8a00-ea66bd76d78b","added_by":"auto","created_at":"2024-05-21 19:09:17","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":435978,"visible":true,"origin":"","legend":"","description":"","filename":"GA.png","url":"https://assets-eu.researchsquare.com/files/rs-4394161/v1/ee90e41feb25606e3f606e58.png"}],"financialInterests":"No competing interests reported.","formattedTitle":"Diagnostic value of magnification optical enhancement by experts and ResNet for detecting diminutive colorectal polyps: A retrospective study","fulltext":[{"header":"Background","content":"\u003cp\u003eColorectal cancer (CRC) is a highly prevalent malignant tumor in clinical practice, with an estimated annual incidence of approximately 1.93\u0026nbsp;million cases worldwide [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. China accounts for a significant proportion of these cases, with an estimated 555 550 new cases reported annually. Fortunately, colonoscopic screening effectively reduces the incidence of CRC [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. However, white-light endoscopy has poor efficacy in differentiating various colonic polyp types [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Hence, an invasive examination should be performed if any abnormality is detected, which can severely increase patient suffering. Such disease types can reportedly be identified in real time according to optical images of electronic chromoendoscopy, such as narrow band imaging (NBI) and optical enhancement (OE), which avoid unnecessary histological examinations of small adenomas and resection of small polyps [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. However, the diagnostic efficacy of inexperienced endoscopists often struggles to achieve thresholds of \u0026lsquo;resect-and-discard\u0026rsquo; and \u0026lsquo;diagnose-and-leave\u0026rsquo; strategies outlined by the Preservation and Incorporation of Valuable Endoscopic Innovations (PIVI) [\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFortunately, recent advances in artificial intelligence (AI) have recognised the potential of computer-assisted algorithms for identifying images, which have been widely used to detect diabetic retinopathy and skin cancer [\u003cspan additionalcitationids=\"CR10\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Computer-assisted algorithms can help endoscopists make accurate diagnoses by providing real-time feedback and guidance and can overcome the limitations of high interobserver variability in polyp identification, leading to better patient outcomes and reduced healthcare costs [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. However, there is a lack of deep learning models for predicting colorectal polyp pathological types during i-scan OE endoscopy, leading to large differences in the prediction of polyp pathological types among different endoscopists [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Therefore, the PIVI strategy becomes difficult to implement, resulting in increased medical expenses.\u003c/p\u003e \u003cp\u003eHerein, we aimed to evaluate the accuracy of i-scan OE with magnification endoscopy for real-time differentiation of colorectal polyp pathological types and assessed whether it can help experts meet \u0026lsquo;resect-and-discard\u0026rsquo; and \u0026lsquo;detect-and-leave\u0026rsquo; strategy requirements [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Furthermore, a computer-assisted ResNet algorithm was trained to automatically differentiate between colorectal polyp pathological types. The diagnostic results of ResNet and AlexNet were compared to verify the superiority of ResNet over the conventional convolutional neural network (CNN) in colonoscopic image-based disease diagnosis.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003ePatient recruitment\u003c/h2\u003e\n \u003cp\u003eTo evaluate the accuracy of i-scan OE with magnification endoscopy for real-time differentiation of colorectal polyp pathological types, patients were recruited between August 2016 and February 2018 at Qilu Hospital, Shandong University. The inclusion criteria were patients who underwent colonoscopy under anaesthesia, provided written informed consent, and had at least one polyp measuring\u0026thinsp;\u0026lt;\u0026thinsp;10 mm. Patients with inflammatory bowel disease, CRC, or hereditary polyposis and those with incomplete colonoscopy or inadequate bowel preparation were excluded.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n \u003ch2\u003eReal-time study design and endoscopy procedure\u003c/h2\u003e\n \u003cp\u003eAll videos were captured using a colonoscope equipped with an i-scan OE (EPK-i7010, Pentax, Japan). A MagniView EC-3890FZI magnifying endoscope with a 4\u0026ndash;100 mm focal range and a maximum optical amplification multiple of 136 was used. Endoscopic operations were performed by two endoscopists using a rubber cap (Pentax Distal Rubber Hood OE-A59, Pentax, Japan).\u003c/p\u003e\n \u003cp\u003eAfter bowel cleaning with 2 L polyethylene glycol, the patients underwent colonoscopy under intravenous propofol anaesthesia. For polyp identification, the standard protocol involved initiating the washing process and simultaneously recording a video of the procedure. The endoscopist used magnified mode 1 to predict polyp pathology and simultaneously recorded the videos.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\n \u003ch2\u003eDiagnosis of pathological types\u003c/h2\u003e\n \u003cp\u003eOur preexperimental findings suggested promising diagnostic efficacy of the Japanese NBI Expert Team (JNET) classification in distinguishing between adenomas and hyperplastic polyps when observed under magnification i-scan OE. Therefore, here, we adopted the JNET classification as the standard for differentiation [\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e]. Lesions classified as normal, hyperplastic, or sessile serrated polyps were categorised as type I, characterised by the absence of visible blood vessels on their surface. In cases where the vessels were visible, the calibres of the vessels in the polyps were similar to those in the surrounding normal mucosa. Low-grade intramucosal neoplasia, including intramucosal cancer with low-grade structural atypia, was classified as type IIA, exhibiting dark brown microvessels that are uniform and arranged in a relatively well-ordered reticular pattern.\u003c/p\u003e\n \u003cp\u003eAn experienced endoscopist determined the pathology of each video clip. After observing the polyps, a biopsy was performed to obtain tissue samples for pathological examination. In cases wherein polyps were found in adjacent segments with similar surface structures, only one polyp was biopsied for further pathological examination. Polyp pathological types were categorised in accordance with the Vienna Classification into nontumorous (including inflammatory and hyperplastic polyps) and tumorous (such as tubular adenomas, tubulopapillary adenomas, and tubular papillary adenomas) polyps [\u003cspan class=\"CitationRef\"\u003e15\u003c/span\u003e]. Patients with hamartomas were excluded from this study.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\n \u003ch2\u003eStatistical analysis\u003c/h2\u003e\n \u003cp\u003eStatistical information regarding patient age, sex, polyp location, polyp morphology, polyp size, and final pathological type was collected and analysed. The polyp location was determined on the basis of the anatomical location. The duodenum and jejunum were classified as parts of the ascending colon, the ileum and jejunum were classified as parts of the transverse colon, and the sigmoid colon and rectum were classified as parts of the descending colon. Polyp morphology was classified in accordance with the Paris classification: flat-elevated (Ip), elevated (Is) and flat-depressed (IIs) [\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e\n \u003cp\u003ePolyp size was estimated by endoscopists using biopsy forceps. The distance between the polyp and colon wall was used to evaluate polyp size.\u003c/p\u003e\n \u003cp\u003eTo assess the diagnostic value of polyps measuring\u0026thinsp;\u0026le;\u0026thinsp;5 mm under magnification i-scan OE, we compared the consistency between the optical diagnosis results and pathological results for these polyps. The corresponding negative predictive value (NPV) was determined to evaluate whether the criteria for PIVI were met.\u003c/p\u003e\n \u003cp\u003eTo assess the performance of the diagnostic process, several statistical measures were calculated, including sensitivity, specificity, positive predictive value (PPV), NPV, and accuracy. The statistical software packages NCSS 11 and R (version 3.3.2) were used to perform the calculations.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n \u003ch2\u003ePost hoc diagnosis with ResNet\u003c/h2\u003e\n \u003cp\u003eVideo files of the i-scan OE were first cut into images because the ResNet constructed here only receives pictures as input. Each second of the videos contained 30 pins, five of which were selected at equal intervals. The images contained three parts, labelled regions A (tissue district), B (edge district) and C (supplementary information district) (Fig. 1(a)). Because only region A contained useful information for classification, the other two regions were removed before further operations. The after-tailored pictures have a pixel dimension of 932 \u0026times; 702, which is sufficient to contain pathological and partially normal tissues (Fig. 1(b)). However, many pictures were repetitive because of the irregular and slow camera movement; therefore, a process of deduplication was needed. All pictures were resised to 10 \u0026times; 10 pixels according to the regional means. The D-hash was obtained by comparing the red-channel values of the pixels with those of adjacent pixels. The result was 0 if the right one was larger than the left one; otherwise, it was 1. Adjacent pictures with the same D-hash sequence were deemed the same, and only one picture was reserved. Finally, the blurry images were artificially removed by two experienced endoscopists to further guarantee sample quality. Typical neoplastic polyp and hyperplastic polyp samples are illustrated in Fig. 1(c) and (d), respectively.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003eConstruction of the deep learning model\u003c/h2\u003e\n \u003cp\u003eAn 18-layer ResNet was adopted as the classification model; its framework is shown in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. Extending the conventional CNN, a mass of residual blocks is contained in ResNet, whose input can be transmitted to its output through a branch consisting only of a one-dimensional convolution layer. ResNet partially overcomes the problem of a disappearing gradient in a deep CNN and possesses a stronger fitting and generalisation ability. Notably, three-channel RGB images were chosen as the input to the network to preserve the most information. Although the complexity of the network was greater than that of the grayscale image, there has been significant progress in diagnostic accuracy. The image dataset was randomly divided into training and test sets at a ratio of 8:2. Weighted sampling was used to mitigate the effects of sample imbalance. For ResNet, the key parameters included the size of the convolution kernel and the mapping size of each layer. Particle swarm optimisation was introduced to optimise these parameters. To simplify the calculations, the parameters of layers other than the first layer were set at a constant ratio; therefore, there were only three parameters to be optimised: the kernel size of the first layer, the kernel size of other layers, and the padding size. The optimisation goal was to maximise the average loss on the test set after training for 100 epochs; the training time was simultaneously used as an important item. The ultimate objective function is expressed as follows:\u003c/p\u003e\n \u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"472\" height=\"59\"\u003e\u003c/p\u003e\n \u003cp\u003ewhere \u003cem\u003er\u003c/em\u003e is the reward, and a higher value indicates a more effective parameter. The \u003cem\u003eAUC\u003c/em\u003e denotes the area under the receiver operating characteristic (ROC) curve of the diagnostic model after training; \u003cem\u003etrain\u003c/em\u003e is the training time in hours for 100 epochs, and b is the bias, which was set to 10 in our experiment.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003ePatients and polyps\u003c/h2\u003e \u003cp\u003eBetween March 2017 and August 2017, a dataset comprising 88 polyps from 58 patients (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) was collected. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarises the polyp characteristics.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCharacteristics of the 88 polyps in the dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e Total (n\u0026thinsp;=\u0026thinsp;88)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNeoplastic (n\u0026thinsp;=\u0026thinsp;51)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHyperplastic (n\u0026thinsp;=\u0026thinsp;37)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSize\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026le;\u0026thinsp;5 mm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e33\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6\u0026ndash;10 mm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLocation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c5\" namest=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRectum\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSigmoid Colon\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescending Colon\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTransverse Colon\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAscending Colon\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCecum\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMorphology\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIs +\u0026thinsp;Ip\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e22\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIIa\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePathology\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e37\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTubular Adenoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTubulovillous Adenoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eDiagnostic performance of experienced endoscopists\u003c/h2\u003e \u003cp\u003eRegarding the diagnostic performance of the deep neural network and endoscopists in differentiating between neoplastic and hyperplastic colorectal polyps (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), experienced endoscopists exhibited 89.77% accuracy, 96% sensitivity for adenoma identification, 81.58% specificity, 93.94% NPV, and 87.27% PPV during radiographic recording for the 88 polyps.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDiagnostic performance of the deep neural network and endoscopists in differentiating colorectal polyp types\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSensitivity, n (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSpecificity, n (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAccuracy, n (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePPV, n (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNPV, n (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"6\" nameend=\"c6\" namest=\"c1\"\u003e \u003cp\u003eRealtime Diagnosis\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExperienced Expert 1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e48/50\u003c/p\u003e \u003cp\u003e(96.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e31/38\u003c/p\u003e \u003cp\u003e(81.58)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e79/88\u003c/p\u003e \u003cp\u003e(89.77)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e48/55\u003c/p\u003e \u003cp\u003e(87.27)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e31/33\u003c/p\u003e \u003cp\u003e(93.94)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"6\" nameend=\"c6\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePost hoc Diagnosis\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExpert 1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e77/89\u003c/p\u003e \u003cp\u003e(86.52)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5/11\u003c/p\u003e \u003cp\u003e(45.45)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e82/100\u003c/p\u003e \u003cp\u003e(82.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e77/83\u003c/p\u003e \u003cp\u003e(92.77)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5/17\u003c/p\u003e \u003cp\u003e(29.41)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExpert 2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e75/89\u003c/p\u003e \u003cp\u003e(84.27)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7/11\u003c/p\u003e \u003cp\u003e(63.64)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e82/100\u003c/p\u003e \u003cp\u003e(82.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e75/79\u003c/p\u003e \u003cp\u003e(94.94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e7/21\u003c/p\u003e \u003cp\u003e(33.33)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExperienced Expert 2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e81/89\u003c/p\u003e \u003cp\u003e(91.01)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8/11\u003c/p\u003e \u003cp\u003e(72.73)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e89/100\u003c/p\u003e \u003cp\u003e(89.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e81/84\u003c/p\u003e \u003cp\u003e(96.43)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e8/16\u003c/p\u003e \u003cp\u003e(50.00)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e83/89\u003c/p\u003e \u003cp\u003e(93.26)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8/11\u003c/p\u003e \u003cp\u003e(72.73)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e91/100\u003c/p\u003e \u003cp\u003e(91.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e83/86\u003c/p\u003e \u003cp\u003e(96.51)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e8/14\u003c/p\u003e \u003cp\u003e(57.14)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExpert 1 with ResNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e81/89\u003c/p\u003e \u003cp\u003e(91.01)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7/11\u003c/p\u003e \u003cp\u003e(63.64)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e88/100\u003c/p\u003e \u003cp\u003e(88.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e81/85\u003c/p\u003e \u003cp\u003e(95.29)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e7/15\u003c/p\u003e \u003cp\u003e(46.67)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExpert 2 with ResNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e77/89\u003c/p\u003e \u003cp\u003e(86.52)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e9/11\u003c/p\u003e \u003cp\u003e(81.82)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e86/100\u003c/p\u003e \u003cp\u003e(86.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e77/79\u003c/p\u003e \u003cp\u003e(97.47)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e9/21\u003c/p\u003e \u003cp\u003e(42.86)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eExperts 1 and 2 were inexperienced endoscopists, whereas Experienced Experts 1 and 2 were experienced endoscopists.\u003c/p\u003e \u003c/div\u003e \n\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\n \u003ch2\u003eImage analysis by the trained ResNet model\u003c/h2\u003e\n \u003cp\u003eThe entire training process, comprising 1000 epochs, took approximately 9.5 hours. By considering adenomatous polyps as positive examples and hyperplastic polyps as negative examples, curves for the accuracy, sensitivity, specificity and NPV of the testing data during training are depicted in Fig. 3(a), (b), (c), and (d), respectively. The accuracy reached a stable value of approximately 95%, with a corresponding sensitivity, specificity and NPV of 97.5%, 94% and 91%, respectively. Thus, the proposed method demonstrates favourable training performance, achieving a balance between sensitivity and specificity. Fig. 4 presents the ROC curves for distinguishing hyperplastic polyps from adenomas using the CNN model. The AUC was 0.98, which indicates that our model has good diagnostic performance in accurately identifying and differentiating colorectal polyps.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\n \u003ch2\u003eComparison between ResNet and endoscopists\u003c/h2\u003e\n \u003cp\u003eA randomly selected set of 100 images was prepared from a dataset outside the training and testing sets, consisting of both hyperplastic polyps and adenomas. This study evaluated the diagnostic capabilities of ResNet and experienced endoscopists. Two inexperienced endoscopists and one experienced endoscopist observed images of colorectal polyps and differentiated adenomas from hyperplastic polyps. In parallel, the trained ResNet model predicted polyp histological type. Five days later, the two experienced endoscopists differentiated polyps again with the help of the ResNet model.\u003c/p\u003e\n \u003cp\u003eThe diagnostic performance of ResNet was comparable to that of experienced endoscopists (Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e). With AI assistance, inexperienced endoscopists demonstrated similar abilities in differentiating between the two histological polyp types.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\n \u003ch2\u003eComparison between ResNet and other networks\u003c/h2\u003e\n \u003cp\u003eTo further confirm the advantages of our diagnostic neural network, we implemented an AlexNet model to perform the same task. As a traditional CNN model, AlexNet is a well-known deep learning model in the field of computer vision and is often used for disease diagnosis. The accuracy of the AlexNet model is shown in Additional File 1(a). The final accuracy achieved on the test set was approximately 91.5%, which required approximately 1000 epochs to reach. Additionally, as shown in Additional File 1(a), the accuracy of AlexNet on the training set reached 100% after 800 epochs, indicating that overfitting occurred in the model and that the effects were severely restrained. The sensitivity and specificity of the AlexNet model are presented in Additional File 1(b) and Additional File 1(c), respectively. In summary, the ResNet-based diagnostic model has advantages over conventional models in terms of its diagnostic effect and generalizability. Thus, our diagnostic neural network outperformed the conventional AlexNet model in terms of accuracy and convergence speed.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, involving 58 patients with colorectal polyps, an experienced endoscopist made real-time judgments regarding colorectal polyp pathological types. The i-scan OE optical diagnosis reached the PIVI threshold for detecting adenomas, achieving a high NPV of 93.94%. To address the low consistency of diagnoses made by different endoscopists, we developed a ResNet-based deep learning model to help doctors improve their diagnostic efficiency. The model showed excellent performance in distinguishing between hyperplastic and neoplastic polyps, with an accuracy of 95%, sensitivity of 97.5%, specificity of 94% and NPV of 91%, and it displayed a performance comparable to that of the diagnosis made by experienced endoscopists. Additionally, the model outperformed other deep learning models, such as AlexNet. Finally, we extracted another 100 images to test the model's ability. The overall accuracy rate was as high as 91%, which is comparable to that of an experienced endoscopist.\u003c/p\u003e \u003cp\u003eTo reduce the unnecessary need for pathological examination of colorectal polyps, two treatment strategies for PIVI have been proposed: \u0026lsquo;resect-and-discard\u0026rsquo; and \u0026lsquo;detect-and-leave\u0026rsquo;. The diagnostic efficacy of PIVI can be met by experienced expert physicians in large medical centres; however, despite the help of electronic chromatic imaging, such as NBI, the results from primary care institutions and nonexpert endoscopists have been unsatisfactory [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. The rapid development of machine learning has provided solutions for this problem.\u003c/p\u003e \u003cp\u003eRecent studies have demonstrated that CNNs are effective in NBI-guided colorectal polyp identification. In 2017 and 2018, two studies were published, both of which showed significant progress in the use of deep learning models to identify colorectal polyp histopathology during NBI and achieved high sensitivity and accuracy [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. With the advancement of deep learning models and new electronic endoscopic technologies, several recent studies have used various deep learning models to accurately differentiate among colorectal polyp pathological types [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. In one study, researchers investigated the effectiveness of a CNN-based deep learning model in assisting endoscopists in differentiating colorectal polyp pathological types using the ELUXEO 7000 endoscopy platform [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. The NPV of AI-assisted optical diagnosis for diminutive rectosigmoid polyps was 91.0%, meeting the required PIVI thresholds. Thus, AI can provide real-time guidance to endoscopists for the differential diagnosis of colorectal polyps based on endoscopic videos.\u003c/p\u003e \u003cp\u003eThe i-scan OE is a novel electronic chromoendoscopy system developed by the Pentax Corporation that incorporates a continuous light-wave enhancement filter in mode 1, which enhances the baseline projection and produces results similar to those of NBI. Although i-scan OE is widely used in many regions, few studies have been conducted on the potential of endoscopists and deep learning models to predict colorectal polyp pathological types using the i-scan OE system.\u003c/p\u003e \u003cp\u003eOur model demonstrated a comparable differentiation ability for colorectal polyp pathological types in a retrospective analysis within approximately 1000 epochs of training, which is faster than that reported in previous research. Moreover, the sensitivity, specificity, accuracy and NPV of our model were similar to those of the aforementioned studies based on electron chromatic images.\u003c/p\u003e \u003cp\u003eThe rapid development of AI has led to the emergence of new deep learning algorithms suitable for image analysis, providing more choices. Herein, a new deep learning algorithm, ResNet, which has never been applied to colorectal polyp differentiation, was adopted. The ResNet model outperformed the AlexNet model, a traditional CNN model, in terms of diagnostic efficacy. After training on a training set of 480 images, ResNet effectively distinguished neoplastic polyps from hyperplastic polyps in the test set. This suggests that in the field of intelligent image classification for endoscopic images, we should follow the emergence of new deep learning models and continuously optimise existing image recognition systems.\u003c/p\u003e \u003cp\u003eOur study is the first to evaluate the diagnostic efficacy of colorectal polyp histology using experts and a deep learning model. We are also among the first to apply the relatively less commonly used deep learning model ResNet to the field of colorectal polyp image identification by training an efficient ResNet model that can accurately identify polyp images. The results of the human\u0026ndash;machine comparison show that the performance of two inexperienced endoscopists improved significantly upon the use of this model, demonstrating that our deep learning model can help improve the efficiency of endoscopic diagnosis after a short period of training.\u003c/p\u003e \u003cp\u003eHowever, our study has several limitations. First, magnifying i-scan OE is not commonly used in colorectal endoscopy because transparent caps increase the difficulty of insertion. Only patients who underwent colonoscopy with a magnified endoscope were included here; therefore, polyps in these patients may not be representative of all patients with polyps. Moreover, we collected information from only 58 patients with colorectal polyps at Qilu Hospital. The collection of additional clinical data from multiple medical centers could validate the effectiveness of our model. Finally, the NPV and proportion of neoplastic polyps included in the study were related to some extent. Patient selection bias may also affect the NPV; therefore, further large-scale randomised controlled trials are needed to validate the diagnostic efficacy of magnifying i-scan OE for predicting the pathological type of colorectal polyps.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eWe developed a deep learning model for the diagnosis of colorectal polyp pathological type based on i-scan OE images with an NPV of 91%, which met the PIVI threshold. Further validation of the ability of this model in a multicentre prospective clinical trial is needed to enable its application to assist endoscopists, especially those who are untrained, in achieving the clinical practice requirements proposed by the PIVI within a short period, thereby changing the current situation wherein the nonselective resection of small colorectal polyps wastes a significant amount of medical resources and causes a significant economic burden for patients.\u003c/p\u003e"},{"header":"List of abbreviations","content":"\u003cp\u003eCRC, colorectal cancer; NBI, narrow band imaging; OE, optical enhancement; PIVI, preservation and incorporation of valuable endoscopic innovations; AI, artificial intelligence; CNN, convolutional neural network; JNET, Japanese NBI Expert Team; NPV, negative predictive value; PPV, positive predictive value; ROC, receiver operating characteristic; AUC, area under the curve\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe current study was approved by the Medical Ethics Committee of Qilu Hospital of Shandong University (no. 2016043) and has completed the review process for registration on a clinical trials database, which can be accessed at https://clinicaltrials.gov/ with the unique identifier NCT02929641. All necessity for written informed consent from patients was waived by the Medical Ethics Committee of Qilu Hospital of Shandong University (no. 2016043 ), as long as patient data remained anonymous. The study adheres to the ethical principles outlined in the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication:\u003c/strong\u003e Not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials:\u0026nbsp;\u003c/strong\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no conflicts of interest for this article.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by the National Natural Science Foundation of China (82270580 and 82070552), the Key Research and Development Program of Shandong Province (2021CXGC010506) and the Taishan Scholars Program of Shandong Province.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; contributions:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eXW: conceptualisation and design; analysis and interpretation of the data; drafting of the article; and final approval of the article\u003c/p\u003e\n\u003cp\u003eYZ: conceptualisation and design; analysis and interpretation of the data; and final approval of the article\u003c/p\u003e\n\u003cp\u003eHY: analysis and interpretation of the data; critical revision of the article for intellectual content; and final approval of the article\u003c/p\u003e\n\u003cp\u003eYL: conceptualisation and design; analysis and interpretation of the data; critical revision of the article for intellectual content;\u0026nbsp;and\u0026nbsp;final approval of the article\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements:\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSome of the graphic elements in the graphical abstract of this article were sourced from FigDraw (ID: USOISb27bb).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2021;71:209\u0026ndash;49.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBretthauer M, L\u0026oslash;berg M, Wieszczy P, Kalager M, Emilsson L, Garborg K, et al. Effect of colonoscopy screening on risks of colorectal cancer and related death. N Engl J Med. 2022;387:1547\u0026ndash;56.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGlissen Brown JR, Mansour NM, Wang P, Chuchuca MA, Minchenberg SB, Chandnani M, et al. Deep learning computer-aided polyp detection reduces adenoma miss rate: a United States multi-center randomized tandem colonoscopy study (CADeT-CS Trial). Clin Gastroenterol Hepatol. 2022;20:1499\u0026ndash;e15074.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVișovan II, Tanțău M, Pascu O, Ciobanu L, Tanțău A. The role of narrow band imaging in colorectal polyp detection. Bosn J Basic Med Sci. 2017;17:152\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIacucci M, Trovato C, Daperno M, Akinola O, Greenwald D, Gross SA, et al. Development and validation of the SIMPLE endoscopic classification of diminutive and small colorectal polyps. Endoscopy. 2018;50:779\u0026ndash;89.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRees CJ, Rajasekhar PT, Wilson A, Close H, Rutter MD, Saunders BP, et al. Narrow band imaging optical diagnosis of small colorectal polyps in routine clinical practice: the Detect Inspect Characterise Resect and Discard 2 (DISCARD 2) study. Gut. 2017;66:887\u0026ndash;95.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKlenske E, Zopf S, Neufert C, N\u0026auml;gel A, Siebler J, Gschossmann J, et al. I-scan optical enhancement for the in vivo prediction of diminutive colorectal polyp histology: results from a prospective three-phased multicentre trial. PLoS ONE. 2018;13:e0197520.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRex DK, Kahi C, O'Brien M, Levin TR, Pohl H, Rastogi A, et al. The American Society for Gastrointestinal Endoscopy PIVI (Preservation and Incorporation of Valuable Endoscopic Innovations) on real-time endoscopic assessment of the histology of diminutive colorectal polyps. Gastrointest Endosc. 2011;73:419\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEsteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542:115\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMookiah MR, Acharya UR, Chua CK, Lim CM, Ng EY, Laude A. Computer-aided diagnosis of diabetic retinopathy: a review. Comput Biol Med. 2013;43:2136\u0026ndash;55.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTakiddin A, Schneider J, Yang Y, Abd-Alrazaq A, Househ M. Artificial intelligence for skin cancer detection: scoping review. J Med Internet Res. 2021;23:e22934.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMin JK, Kwak MS, Cha JM. Overview of deep learning in gastrointestinal endoscopy. Gut Liver. 2019;13:388\u0026ndash;93.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIacucci M, Gasia M, Akinola O, et al. Mo1743 accuracy and inter-observer agreement for characterization of colonic polyps using the new optical enhancement -iSCAN Closefocus\u0026trade; Colonoscope. Gastroenterology. 2016;150:S769.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShaukat A, Kaltenbach T, Dominitz JA, Robertson DJ, Anderson JC, Cruise M, et al. Endoscopic recognition and management strategies for malignant colorectal polyps: recommendations of the us multi-society task force on colorectal cancer. Gastroenterology. 2020;159:1916\u0026ndash;e19342.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchlemper RJ, Riddell RH, Kato Y, Borchard F, Cooper HS, Dawsey SM, et al. The Vienna classification of gastrointestinal epithelial neoplasia. Gut. 2000;47:251\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eByrne MF, Chapados N, Soudan F, Oertel C, Linares P\u0026eacute;rez M, Kelly R, et al. Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during analysis of unaltered videos of standard colonoscopy using a deep learning model. Gut. 2019;68:94\u0026ndash;100.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen PJ, Lin MC, Lai MJ, Lin JC, Lu HH, Tseng VS. Accurate classification of diminutive colorectal polyps using computer-aided analysis. Gastroenterology. 2018;154:568\u0026ndash;75.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGong EJ, Bang CS, Lee JJ, Seo SI, Yang YJ, Baik GH, et al. No-code platform-based deep-learning models for prediction of colorectal polyp histology from white-light endoscopy images: development and performance verification. J Pers Med. 2022;12:963.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePacal I, Karaboga D. A robust real-time deep learning based automatic polyp detection system. Comput Biol Med. 2021;134:104519.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRondonotti E, Hassan C, Tamanini G, Antonelli G, Andrisani G, Leonetti G, et al. Artificial intelligence-assisted optical diagnosis for the resect-and-discard strategy in clinical practice: the Artificial intelligence BLI Characterization (ABC) study. Endoscopy. 2023;55:14\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"i-scan OE, ResNet, colorectal polyps, pathological types","lastPublishedDoi":"10.21203/rs.3.rs-4394161/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4394161/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eDeep learning models, such as narrow band imaging, are useful for identifying colorectal polyp pathological types using electronic chromoendoscopy. However, such findings remain unreported for i-scan optical enhancement (OE). Thus, this study proposes a ResNet-based model to distinguish between colorectal polyp types.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eEndoscopic videos of 58 patients with colorectal polyps were collected, and their pathological types were determined by pathological examination. An experienced endoscopist was asked to predict the polyp pathological types shown in the videos to evaluate the diagnostic accuracy of the endoscopists. The same videos were extracted as images to train the ResNet model and evaluate its effectiveness. Finally, 100 endoscopic images were randomly selected for human\u0026ndash;machine comparisons to examine the consistency between machine learning and endoscopist diagnoses.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eAn accuracy of 89.77% was obtained by experienced endoscopists in distinguishing colorectal polyp types based on the entire video; the accuracy of the ResNet model based on video-extracted images was approximately 95%. Therefore, the ability of the ResNet model to classify colorectal polyps matched that of an experienced endoscopist.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eI-scan OE provides a powerful tool for experienced endoscopists to differentiate between colorectal polyp pathological types; the ResNet model can help endoscopists achieve high diagnostic accuracy and reduce unnecessary invasive examinations.\u003c/p\u003e","manuscriptTitle":"Diagnostic value of magnification optical enhancement by experts and ResNet for detecting diminutive colorectal polyps: A retrospective study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-05-21 19:01:12","doi":"10.21203/rs.3.rs-4394161/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-11-26T11:45:01+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-26T03:49:44+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"49857643738749712636260556623423677820","date":"2025-11-10T15:10:03+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-03-12T18:45:00+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"255809300841834123011971733234515112003","date":"2025-02-24T15:06:49+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"184881670736298297644854500416793509823","date":"2025-02-16T13:50:52+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"126356292111194107642635611406598827816","date":"2024-12-16T12:25:23+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-07-15T09:11:56+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2024-05-10T04:05:50+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-05-10T04:02:00+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-05-10T04:02:00+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2024-05-09T09:34:14+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"2261b34d-5980-4cf0-ae32-f65eadaa2be6","owner":[],"postedDate":"May 21st, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[],"tags":[],"updatedAt":"2026-03-19T11:10:20+00:00","versionOfRecord":[],"versionCreatedAt":"2024-05-21 19:01:12","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4394161","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4394161","identity":"rs-4394161","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}