Development and validation of a real-time AI model for differentiating benign and malignant gastric ulcers: A multicenter retrospective study

Development and validation of a real-time AI model for differentiating benign and malignant gastric ulcers: A multicenter 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 Development and validation of a real-time AI model for differentiating benign and malignant gastric ulcers: A multicenter retrospective study Yibo Tan, Yongjun Wu, Junyu Lu, Song He, Zhihang Zhou This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8776954/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 9 You are reading this latest preprint version Abstract Aim: To develop and validate a deep learning-based AI system for the dynamic, real-time differentiation of benign and malignant gastric ulcers during endoscopy, with the goal of enhancing diagnostic precision and reducing superfluous biopsies. Methods: This was a multicenter, retrospective study collecting endoscopic images and videos from four tertiary hospitals in China. An improved YOLOv8 model, incorporating an illumination attention module, was developed for real-time instance segmentation and classification. The dataset comprised 9,840 benign ulcer images, 1,727 malignant ulcer images, and 15,791 normal mucosa images, split into training, testing, and validation sets at an 8:1:1 ratio. Performance was evaluated based on precision, recall, specificity, and processing latency. Results: On the validation set, the AI model achieved an overall precision, recall, and specificity of 0.91, 0.91, and 0.95, respectively. For malignant ulcer recognition specifically, the precision, recall, and specificity were 0.90, 0.91, and 0.99. The model demonstrated strong real-time performance with a latency of 8.84 ms per frame and a processing speed of 113 frames per second. Conclusion: The developed AI model enables accurate, real-time discrimination between benign and malignant gastric ulcers during endoscopy. It holds significant potential to augment clinical decision-making, standardize diagnostic quality, and optimize biopsy strategies. Artificial Intelligence Gastric Ulcer Benign and Malignant Differentiation Deep Learning Endoscopic Diagnosis Figures Figure 1 Figure 2 Figure 3 Figure 4 1 Introduction Gastric ulcer, one of the more common types of peptic ulcers, refers to a condition in which the gastric mucosa is eroded by gastric acid and pepsin, resulting in a breach that extends beyond the muscularis mucosae 1 .As a prevalent disease of the digestive system, clinical emphasis is placed on distinguishing between benign and malignant forms of gastric ulcer. Malignant gastric ulcer, also referred to as ulcerative gastric cancer, represents the most frequent pathological manifestation of gastric cancer. Gastric cancer ranks among the top five most common malignant cancers globally and is associated with a high mortality rate. The survival rate for gastric cancer in most regions of the world is approximately 20%, whereas the five-year survival rate for early-stage gastric cancer can reach 90% 2–4 .Achieving early and accurate diagnosis of ulcerative gastric cancer poses significant challenges. For instance, there is considerable overlap in symptoms and endoscopic findings between benign ulcers and early gastric cancer. Additionally, following treatment with potent acid-suppressive medications, gastric cancer patients may experience relief from abdominal pain, and the ulcer may shrink or partially heal 5 , 6 . Endoscopy with biopsy remains the gold standard for the diagnosis and differential diagnosis of gastric ulcers. However, it has certain limitations:First, the endoscopic assessment of ulcer nature is subjective and heavily reliant on the operator's experience. Endoscopists primarily diagnose malignant gastric ulcers based on morphological features, such as an irregular, nodular ulcer base, atypical shape, foul and thick exudate, and rough or raised margins. Nevertheless, some benign ulcers with a long history can also exhibit macroscopic features characteristic of malignancy under endoscopy 7 ,and one study demonstrated that the sensitivity of endoscopy for diagnosing gastric ulcers is 92% 8 .Furthermore, a significant issue is the high miss rate associated with endoscopic examination. Several studies have reported that the proportion of undetected gastric cancers reaches 9.8% to 25.8%. A detailed analysis indicated that 73% of these missed cases were attributable to endoscopist error 9 , 10 .Second, while guidelines recommend obtaining biopsies from lesions with high-risk features identified under white-light imaging, the biopsy rates among endoscopists vary considerably, ranging from 22.4% to 52.9%. Concurrently, as the number of biopsies increases, so does the number of histologically negative biopsies from endoscopically normal-appearing mucosa, imposing an unnecessary burden on both patients and healthcare systems 11 .Third, the accuracy of biopsy is dependent on the number and location of the samples obtained. Endoscopic biopsy is an invasive procedure that can cause mucosal damage and bleeding. Repeated biopsies or taking multiple samples over a large area may even lead to submucosal fibrosis 12 , 13 . We aim to enhance the real-time dynamic endoscopic diagnostic accuracy for ulcer characterization and reduce unnecessary biopsies. Emerging technologies offer promising avenues to achieve this goal. Artificial intelligence (AI) has demonstrated considerable potential in auxiliary diagnosis: in recent years, AI models based on deep learning (DL) have made significant strides in medical image analysis. Current studies have indicated the applicability of AI models in automatically assessing the malignant potential of gastric ulcer images. However, most of these studies rely on single-center data and focus on classifying gastric conditions (e.g:normal mucosa, benign ulcers, malignant ulcers) based on individual static images. This approach diverges from the clinical reality, where endoscopists make diagnoses through continuous observation and holistic assessment of the entire lesion 14 – 17 . Therefore, we plan to develop an AI system using multi-center data to dynamically identify gastric ulcer lesions in real-time during endoscopy. This system will perform benign versus malignant classification and provide pixel-level delineation of the lesions, thereby improving the accuracy of ulcer identification and offering guidance on the necessity of biopsy. 2 Materials and Methods 2.1 Study design A brief summary of the study design is shown in Fig. 1 .This is a multicenter, retrospective study. All endoscopic images and videos used in this research were retrospectively collected from patients who underwent gastroscopy during routine clinical practice between January 1, 2021, and April 30, 2025, at four tertiary care hospitals in China. The participating hospitals are as follows: The Second Affiliated Hospital of Chongqing Medical University (Chongqing, China), Chongqing University Three Gorges Hospital (Chongqing, China), Chongqing Kaizhou District People's Hospital (Chongqing, China), and Chengdu Third People's Hospital (Chengdu, China). All images and videos were captured using standard endoscopes (GIF-Q260, GIF-H260, GIF-H290, Olympus Medical Systems Co. Ltd and HD-500,HD-550, SonoScape Biomedical Technology Co., Ltd). The study was approved by the respective Institutional Review Boards and was conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived by the Institutional Review Boards due to the retrospective nature of the study and the use of fully anonymized data. 2.2 Datasets and preprocessing The inclusion criteria for images were as follows: 1) Patients aged over 18 who underwent gastroscopy and were histologically diagnosed with benign or malignant ulcerative lesions according to the World Health Organization (WHO) classification; 2) Endoscopic images without ulcerative lesions were included as the "normal gastric mucosa" dataset. The exclusion criteria were: 1) Cases where pathological diagnosis recommended further investigations (such as immunohistochemistry) for definitive characterization, but the patients did not complete these tests; 2) Endoscopic image quality insufficient for reliable endoscopic diagnosis; 3) Inadequate pathological diagnosis due to issues in tissue sampling, sectioning, or staining. After selection, a total of 2,028 subjects were included. Additionally, 11 gastroscopy videos from the Second Affiliated Hospital of Chongqing Medical University were obtained, and images were extracted from the real-time video streams. This process ultimately yielded 9,840 benign ulcer images, 1,727 malignant ulcer images, and 15,791 normal background images. All images were divided into training, testing, and validation sets in an 8:1:1 ratio. These representative images were used for model development and validation.A detailed summary of the number and distribution of each category in the datasets is shown in the Supplementary table 1 and 2 .Three physicians, each with over five years of experience in endoscopic examination, annotated the boundaries of each lesion in the images. Two additional endoscopists with more than ten years of endoscopic experience reviewed all annotated images. Any disagreements were resolved through discussion until a consensus was reached. Pathological results were reviewed by experienced pathologists specializing in gastric abnormalities, each possessing over ten years of diagnostic experience 18 .For image preprocessing, an object detection model was trained to automatically extract the Region of Interest (ROI) from the output images. The overall procedure is illustrated in Fig. 2 . All final images were output in JPG format. 2.3 Building of AI Model A lightweight You Only Look Once Version 8 (YOLOv8) model was employed to construct a real-time gastric ulcer instance segmentation system, which simultaneously classifies the malignancy status of gastric ulcers and segments the ulcer regions. To address the impact of varying illumination conditions, a customized Illumination-Aware Attention module was integrated into the original YOLOv8 architecture. The model was developed using the Python programming language and the PyTorch framework, and training was conducted on three NVIDIA GeForce RTX 3080 GPUs (each with 24GB memory). More detailed information regarding the methodologies is provided in the Figure S1 and S2. 2.4 Model evaluation and Statistical methods 2.4.1 Classification performance evaluation Based on the test set, an N×N matrix (N = 3: normal mucosa, benign ulcer, malignant ulcer) was constructed. The rows of the matrix represent the categories marked by the pathological gold standard, and the columns represent the categories predicted by the model. The corresponding precision, specificity and recall were calculated. 2.4.2 Real-time evaluation Latency per Frame: The time required to process a single image from input to output(ms). 3 Results 3.1 When the AI model detects gastric ulcer lesions from the input data of the test image, the AI model will output the disease name (benign ulcer or malignant ulcer) and its probability score (range 0–1). The higher the probability score, the higher the confidence in the diagnosis of the model, and the detected lesions will be outlined on the image. Meanwhile, the heat map provided by Gradient-weighted class activation mapping(Grad-CAM) indicates that our model makes reasonable and objective judgments based on specific parts of the image, similar to how endoscopists diagnose lesions. Representative cases are given in Fig. 3 and Figure S3 . 3.2 The actual number of pictures in the validation set was 2713, among which the number of lesions of normal gastric ulcer, benign gastric ulcer and malignant gastric ulcer was 1482, 1133 and 154 respectively. Table 1 shows that the precision rate, recall and specificity of artificial intelligence in lesion recognition are 0.91, 0.91and 0.95, among which the precision rate, recall and specificity of benign lesion recognition are 0.83, 0.94 and 0.89 and the precision rate, recall and specificity of malignant lesion recognition are all 0.90, 0.91 and 0.99.The more detailed data are presented in Fig. 4. Table 1 A summary of detection performance of the AI model Precision Recall Specificity Benign ulcer 0.83 0.94 0.89 Malignant ulcer 0.90 0.91 0.99 Background 0.96 0.88 0.95 All 0.91 0.91 0.95 3.3 To evaluate the clinical real-time performance of the model, we recorded the throughput on the validation set. Table 2 shows the speed of artificial intelligence in lesion recognition and processing. Table 2 A summary of real-time performance of the AI model . Number Time(s) Speed(fps) latency(ms) 2713 24 113 8.84 4 Discussion In this multicenter retrospective study, we developed and validated a deep learning-based artificial intelligence model for the differential diagnosis of benign and malignant gastric ulcers. The model achieved an overall accuracy of 0.91 with a single-frame latency of only 8.84 ms.This research not only provides a novel tool for precise diagnosis of gastric ulcers but also establishes a technical pathway for optimizing clinical biopsy decisions and reducing healthcare costs. The system holds particular significance for less experienced endoscopists and challenging ulcer lesions, demonstrating potential to enhance the quality of endoscopic examinations and bridge diagnostic disparities between primary care hospitals and advanced medical centers. Artificial intelligence systems have demonstrated impressive capabilities in medical image analysis. Regarding diagnostic models for gastric ulcer lesions, most previous studies were developed using single-center datasets based on individual static images. However, heterogeneity and diversity across datasets can significantly impact AI performance, making datasets that adequately represent real-world variations crucial 19 , 20 . Our dataset, sourced from four different tertiary hospitals and constructed by extracting frames from real procedure videos, simulates the endoscopist's process of adjusting perspectives and observing dynamic changes in lesions during examination. This design enables the model to capture subtle features often overlooked in static images (such as border motility and adhesion stability of the exudate). For instance, the exudate covering benign ulcers may partially detach after air insufflation, whereas in malignant ulcers, it typically adheres more firmly due to infiltrative growth. Such dynamic characteristics are effectively captured through video frame extraction, reflecting the clinical reasoning process of endoscopists during gastroscopy and thereby enhancing the representativeness of our samples in real-world clinical scenarios.Furthermore, we incorporated Grad-CAM as a post-modeling technique to provide visual explanations. By deconstructing the AI model's diagnostic logic through both post-modeling and pre-modeling approaches, we achieved concretization of abstract theories and improved interpretability of the AI diagnostics, thereby enhancing diagnostic accuracy and consistency 21 . Additionally, some images collected from older endoscopic systems exhibited low brightness. To address the common issue of uneven illumination in endoscopic images 22 , 23 ,this study introduced an illumination-aware attention module into the YOLOv8 architecture. This module dynamically integrates original features with normalized features through feature channel normalization and a differentiable gating mechanism, effectively suppressing interference from reflective areas on the model's judgment. This approach partially mitigates the selection bias often faced in retrospective studies.Ultimately, our AI model achieved an overall diagnostic accuracy of 0.91, which is comparable to the performance of expert endoscopists reported in related literature and superior to that of novice endoscopists 24 . Moreover, in the selection of optimization strategies for medical imaging AI models, the field of digestive endoscopy presents clinical requirements distinctly different from other imaging modalities. Traditional auxiliary diagnostics in imaging—such as radiological image analysis and pathological slide recognition—typically employ offline batch processing modes, with optimization objectives focused on maximizing system throughput to enhance overall processing efficiency. However, in real-time diagnostic and therapeutic scenarios of digestive endoscopy, due to the time-accumulated effects of anesthesia risk and the spatial constraints of endoscopic manipulation 25 , 26 ,clinical decision-making operates within strict time windows—when endoscopists identify a suspicious ulcer lesion, they must complete a series of decisions including determining the nature of the lesion, assessing the necessity for biopsy, and accurately locating it within a short timeframe. This requirement for real-time interaction demands that auxiliary diagnostic models prioritize the optimization of single-frame processing latency. By reducing the inference time per frame, the cumulative effect of operational delays can be avoided, thereby ensuring the continuity of the diagnostic-therapeutic process and the timeliness of decisions.In previous studies, the total time required to complete endoscopy with or without AI assistance has been used as one of the outcome measures. These studies found no statistically significant difference in procedure time between the AI-assisted and manual groups, but this does not fully reflect the real-time performance of the model 27 . The human eye perceives a sequence of images as a continuously moving scene when the frame rate reaches 16–24 frames per second (fps). Therefore, the system must possess efficient image processing capabilities, with a real-time processing efficiency of > 16 frames/second, and the image delay should be < 50 ms compared to the original endoscopic video system. This ensures that physicians can observe and diagnose seamlessly, receiving timely real-time feedback and analysis results, thus fulfilling the requirements for dynamic recognition 28 . Consequently, we recorded the processing time of our AI model for each frame in the data stream, and the results demonstrate that our model exhibits excellent real-time performance. There are some limitations in this study.First, the retrospective design inherently makes it difficult to eliminate selection bias. Second, while the multi-center data source enriches sample diversity, previous research indicates that multi-center data often leads to greater diversity in negative cases but less diversity in positive cases, which may attenuate the identification of positive lesions—a trend consistent with our findings 29 . In the future, we will further optimize the sample size and ratio and employ novel algorithms to overcome this issue. Additionally, all included medical institutions in this study are tertiary hospitals. We should appropriately incorporate case information from primary hospitals in resource-limited areas for model training and validation, which would be more conducive to the future validation and generalization of the AI system. Third, our study is based on an end-to-end algorithm, whose diagnostic process remains an opaque and unexplainable "black box." Although we adopted post-modeling and pre-modeling approaches during model development to enhance medical interpretability, the issue of explainability in AI systems remains not fully resolved, which to some extent affects the credibility and acceptability of the AI system in clinical practice. Fourth, while we treated videos as image sequences and used frame extraction to partially replicate the real examination process, it is essential to validate the model's clinical applicability for real-time diagnosis using datasets containing complete videos. We currently lack prospective experiments to further verify this aspect. In summary, this study successfully developed an artificial intelligence model capable of real-time dynamic differentiation between benign and malignant gastric ulcers by utilizing multicenter data and an improved YOLOv8 architecture. The model demonstrated excellent performance in detecting retrospective images and videos, and despite existing data and technical limitations, it has shown potential to transform traditional endoscopic diagnostic workflows. Future research should focus on multimodal integration, prospective validation, and the establishment of ethical frameworks to promote its comprehensive transition from technological exploration to routine clinical application. Abbreviations AI artificial intelligence DL deep learning WHO World Health Organization ROI Region of Interest YOLOv8 You Only Look Once Version 8 Grad-CAM Gradient-weighted class activation mapping FPS frames per second Declarations Ethics approval and consent to participate The study protocol, including the request for waiver of informed consent, was initially reviewed and approved by the Institutional Review Board (The Ethics Committee of the Second Affiliated Hospital of Chongqing Medical University; Approval No. 2024 Ethical Review No.129). The requirement for informed consent was waived by the Institutional Review Boards due to the retrospective nature of the study and the use of fully anonymized data. Consent for publication Not applicable. Availability of data and materials The datasets used during the current study are available from the corresponding author on reasonable request. Competing Interests The authors declare that they have no competing interests. Funding No funding was received for this research Authors' contributions All authors contributed to the study conception and design.Yibo Tan ：Conceptualization, Data Curation, Statistic Analysis,Writing - Original Draft；YongJun Wu ：Software，Original Draft；Junyu Lu ，Zhihang Zhou，SongHe：Conceptualization, Writing- Reviewing and Editing and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. Acknowledgements Not applicable References Lanas A, Chan FKL. Peptic ulcer disease. Lancet (London, England). Aug. 2017;5(10094):613–24. Smyth EC, Nilsson M, Grabsch HI, van Grieken NC, Lordick F. Gastric cancer. Lancet (London, England). Aug. 2020;29(10251):635–48. Ren J, Jin X, Li J, et al. The global burden of peptic ulcer disease in 204 countries and territories from 1990 to 2019: a systematic analysis for the Global Burden of Disease Study 2019. Int J Epidemiol Oct. 2022;13(5):1666–76. Katai H, Ishikawa T, Akazawa K et al. Jan. Five-year survival analysis of surgically resected gastric cancer cases in Japan: a retrospective analysis of more than 100,000 patients from the nationwide registry of the Japanese Gastric Cancer Association (2001–2007). Gastric cancer: official journal of the International Gastric Cancer Association and the Japanese Gastric Cancer Association. 2018;21(1):144–54. Chinese Medical Association, Chinese Medical Association Journal Press, Chinese Medical Association Digestive Disease Branch, Chinese Medical Association General Practice Branch, Editorial Committee of Chinese Journal of General Practice, Working Group for the Development of Primary Care Guidelines for Digestive System Diseases. Primary Care Guidelines for Peptic Ulcer. (2023). Chinese Journal of General Practice. 2023;22(11):1108–1117. Vakil N. Peptic Ulcer Disease: A Review. Jama Dec. 2024;3(21):1832–42. Gielisse EA, Kuyvenhoven JP. Follow-up endoscopy for benign-appearing gastric ulcers has no additive value in detecting malignancy: It is time to individualise surveillance endoscopy. Gastric cancer: official J Int Gastric Cancer Association Japanese Gastric Cancer Association Oct. 2015;18(4):803–9. Dooley CP, Larson AW, Stace NH, et al. Double-contrast barium meal and upper gastrointestinal endoscopy. A comparative study. Annals of internal medicine. Oct. 1984;101(4):538–45. Menon S, Trudgill N. How commonly is upper gastrointestinal cancer missed at endoscopy? A meta-analysis. Endoscopy Int open Jun. 2014;2(2):E46–50. Yalamarthi S, Witherspoon P, McCole D, Auld CD. Missed diagnoses in patients with upper gastrointestinal cancers. Endoscopy Oct. 2004;36(10):874–9. Januszewicz W, Wieszczy P, Bialek A, et al. Endoscopist biopsy rate as a quality indicator for outpatient gastroscopy: a multicenter cohort study with validation. Gastrointest endoscopy Jun. 2019;89(6):1141–9. Graham DY, Schwartz JT, Cain GD, Gyorkey F. Prospective evaluation of biopsy number in the diagnosis of esophageal and gastric carcinoma. Gastroenterol Feb. 1982;82(2):228–31. Bang CS, Baik GH, Kim JH, et al. Effect of Training in Upper Endoscopic Biopsy. Korean J Helicobacter Up Gastrointest Res. 2015;3(1):33–8. Klang E, Barash Y, Levartovsky A, Barkin Lederer N, Lahat A. Differentiation Between Malignant and Benign Endoscopic Images of Gastric Ulcers Using Deep Learning. Clin Exp Gastroenterol. 2021;14:155–62. Nam JY, Chung HJ, Choi KS, et al. Deep learning model for diagnosing gastric mucosal lesions using endoscopic images: development, validation, and method comparison. Gastrointest endoscopy Feb. 2022;95(2):258–e268210. Namikawa K, Hirasawa T, Nakano K, et al. Artificial intelligence-based diagnostic system classifying gastric cancers and ulcers: comparison between the original and newly developed systems. Endoscopy Dec. 2020;52(12):1077–83. Liu Y, Zhang L, Hao Z, et al. An xception model based on residual attention mechanism for the classification of benign and malignant gastric ulcers. Sci Rep Sep. 2022;13(1):15365. LiZhaoshen. Experts consensus on quality control system of data collection and labeling for artificial intelligence in digestive endoscopy (2019, Shanghai). Chin J Dig Endoscopy.37(8):533–9. Norgeot B, Quer G, Beaulieu-Jones BK, et al. Minimum information about clinical artificial intelligence modeling: the MI-CLAIM checklist. Nat Med Sep. 2020;26(9):1320–4. Ryan M, Stahl BC. Artificial intelligence ethics guidelines for developers and users: clarifying their content and normative implications. J Inform Communication Ethics Soc. 2020;19(1):61–86. Dong Z, Wang J, Li Y, et al. Explainable artificial intelligence incorporated with domain knowledge diagnosing early gastric neoplasms under white light endoscopy. NPJ digital medicine. Apr. 2023;12(1):64. Bi C, Liu B, Wang T et al. Gastric Mucosal Blood Flow Analysis of Ulcer-healing by Electronic Endoscope System.Chin. J Gastroenterol. 2000(04):240–2. Extracting clinical information from endoscopic capsule exams using MPEG-7 visual descriptors. 2nd European Workshop on the Integration of Knowledge, Semantics and Digital Media Technology (EWIMT 2005):105–110. Yuan XL, Zhou Y, Liu W, et al. Artificial intelligence for diagnosing gastric lesions under white-light endoscopy. Surg endoscopy Dec. 2022;36(12):9444–53. Manno M, Deiana S, Gabbani T, et al. Implementation of the European Society of Gastrointestinal Endoscopy (ESGE) and European Society of Gastroenterology and Endoscopy Nurses and Associates (ESGENA) sedation training course in a regular endoscopy unit. Endoscopy Jan. 2021;53(1):65–71. Xu C, Zhu Y, Wu L, et al. Evaluating the effect of an artificial intelligence system on the anesthesia quality control during gastrointestinal endoscopy with sedation: a randomized controlled trial. BMC anesthesiology Oct. 2022;7(1):313. Yuan XL, Liu W, Lin YX, et al. Effect of an artificial intelligence-assisted system on endoscopic diagnosis of superficial oesophageal squamous cell carcinoma and precancerous lesions: a multicentre, tandem, double-blind, randomised controlled trial. The lancet. Gastroenterol Hepatol Jan. 2024;9(1):34–44. Big Data Collaboration Group, Digestive Endoscopology Branch of Chinese Medical Association. Expert consensus on the clinical application of artificial intelligence system to upper gastrointestinal endoscopy (2023, Wuhan). Chin J Dig Endoscopy. 2024;41(2):85–96. Dong Z, Tao X, Du H, et al. Exploring the challenge of early gastric cancer diagnostic AI system face in multiple centers and its potential solutions. J Gastroenterol Oct. 2023;58(10):978–89. Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterials.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 10 Mar, 2026 Reviews received at journal 05 Mar, 2026 Reviews received at journal 03 Mar, 2026 Reviewers agreed at journal 27 Feb, 2026 Reviewers agreed at journal 27 Feb, 2026 Reviewers invited by journal 11 Feb, 2026 Editor assigned by journal 06 Feb, 2026 Submission checks completed at journal 05 Feb, 2026 First submitted to journal 05 Feb, 2026 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-8776954","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":589817325,"identity":"4651b33c-3ef2-4cef-935b-a93a34ced1e2","order_by":0,"name":"Yibo Tan","email":"","orcid":"","institution":"The Second Affiliated Hospital of Chongqing Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yibo","middleName":"","lastName":"Tan","suffix":""},{"id":589817332,"identity":"01b63c45-ec8e-49e4-92f5-274be54c1a7a","order_by":1,"name":"Yongjun Wu","email":"","orcid":"","institution":"Chongqing Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Yongjun","middleName":"","lastName":"Wu","suffix":""},{"id":589817333,"identity":"5a3de31f-e6ec-48ac-96ce-6f809f2e545c","order_by":2,"name":"Junyu Lu","email":"","orcid":"","institution":"The Second Affiliated Hospital of Chongqing Medical University","correspondingAuthor":false,"prefix":"","firstName":"Junyu","middleName":"","lastName":"Lu","suffix":""},{"id":589817334,"identity":"c8aae0a0-5edc-4f2c-b111-0ab5c22e81c7","order_by":3,"name":"Song He","email":"","orcid":"","institution":"The Second Affiliated Hospital of Chongqing Medical University","correspondingAuthor":false,"prefix":"","firstName":"Song","middleName":"","lastName":"He","suffix":""},{"id":589817336,"identity":"10eeeae1-98d3-47c7-90da-7c2c85571a77","order_by":4,"name":"Zhihang Zhou","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8klEQVRIiWNgGAWjYDACCSBmbJBIAFEPoGIGRGthhiklSgsDSAubBFFa5Gc3H5P8ucMij392+7WKH2XbEhvYm7dJMNTcwamFcc6xNAnJMxLFEnfOlN3sOXc7sYHnWJkEw7FnOLUwS+SYSRi2SSQ23MhJu83YBtQCEmFsOIxTCxtIQSJQy3yglmKwFvk3+LXwgLQcBGrZcCP9GDPEFh78WiQk0pItG4FaNt7IYZYE+sW4jSet2CLhGG4t8jOSD9782VaXOO9G+sMPP8puy/azH95440MNbi3IbgRGBxsYMYCjiQjA/gCqfhSMglEwCkYBKgAA4rdXEI4FII4AAAAASUVORK5CYII=","orcid":"","institution":"The Second Affiliated Hospital of Chongqing Medical University","correspondingAuthor":true,"prefix":"","firstName":"Zhihang","middleName":"","lastName":"Zhou","suffix":""}],"badges":[],"createdAt":"2026-02-03 14:11:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8776954/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8776954/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102839443,"identity":"fa6a6061-b5a1-430d-8816-88dd1b312a47","added_by":"auto","created_at":"2026-02-17 11:45:42","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1328645,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u0026nbsp;\u003c/p\u003e","description":"","filename":"figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8776954/v1/41e3da4234d8ee7e3fcb48a7.jpg"},{"id":102839446,"identity":"5ac20ee6-4191-46e0-a870-5dcdd156b7cd","added_by":"auto","created_at":"2026-02-17 11:45:42","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":3022339,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u0026nbsp;\u003c/p\u003e","description":"","filename":"figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8776954/v1/35cc9f8b548c5548553956de.jpg"},{"id":102839448,"identity":"87f4fb00-a8bf-4fda-bf87-8929f4f8cb37","added_by":"auto","created_at":"2026-02-17 11:45:43","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":5506486,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u0026nbsp;\u003c/p\u003e","description":"","filename":"figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8776954/v1/0bb8e588ceb383a3ac289fb5.jpg"},{"id":102839444,"identity":"be4ccc56-6d68-47f7-9378-0515233d42ee","added_by":"auto","created_at":"2026-02-17 11:45:42","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":2041968,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u0026nbsp;\u003c/p\u003e","description":"","filename":"figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8776954/v1/10d397fa1902a3eefe8d619f.jpg"},{"id":103504193,"identity":"0aa00444-fed3-49b4-8631-9a936424e13e","added_by":"auto","created_at":"2026-02-26 13:18:18","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":12451180,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8776954/v1/5276e861-c393-4f20-9e6e-d8f7592f5c3b.pdf"},{"id":102839445,"identity":"773df12e-d783-4095-825f-4b8020053166","added_by":"auto","created_at":"2026-02-17 11:45:42","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1419424,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-8776954/v1/152e320f751439785151760e.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eDevelopment and validation of a real-time AI model for differentiating benign and malignant gastric ulcers: A multicenter retrospective study\u003c/p\u003e","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eGastric ulcer, one of the more common types of peptic ulcers, refers to a condition in which the gastric mucosa is eroded by gastric acid and pepsin, resulting in a breach that extends beyond the muscularis mucosae\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e.As a prevalent disease of the digestive system, clinical emphasis is placed on distinguishing between benign and malignant forms of gastric ulcer. Malignant gastric ulcer, also referred to as ulcerative gastric cancer, represents the most frequent pathological manifestation of gastric cancer. Gastric cancer ranks among the top five most common malignant cancers globally and is associated with a high mortality rate. The survival rate for gastric cancer in most regions of the world is approximately 20%, whereas the five-year survival rate for early-stage gastric cancer can reach 90% \u003csup\u003e2\u0026ndash;4\u003c/sup\u003e.Achieving early and accurate diagnosis of ulcerative gastric cancer poses significant challenges. For instance, there is considerable overlap in symptoms and endoscopic findings between benign ulcers and early gastric cancer. Additionally, following treatment with potent acid-suppressive medications, gastric cancer patients may experience relief from abdominal pain, and the ulcer may shrink or partially heal \u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eEndoscopy with biopsy remains the gold standard for the diagnosis and differential diagnosis of gastric ulcers. However, it has certain limitations:First, the endoscopic assessment of ulcer nature is subjective and heavily reliant on the operator's experience. Endoscopists primarily diagnose malignant gastric ulcers based on morphological features, such as an irregular, nodular ulcer base, atypical shape, foul and thick exudate, and rough or raised margins. Nevertheless, some benign ulcers with a long history can also exhibit macroscopic features characteristic of malignancy under endoscopy \u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e,and one study demonstrated that the sensitivity of endoscopy for diagnosing gastric ulcers is 92% \u003csup\u003e8\u003c/sup\u003e.Furthermore, a significant issue is the high miss rate associated with endoscopic examination. Several studies have reported that the proportion of undetected gastric cancers reaches 9.8% to 25.8%. A detailed analysis indicated that 73% of these missed cases were attributable to endoscopist error \u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e.Second, while guidelines recommend obtaining biopsies from lesions with high-risk features identified under white-light imaging, the biopsy rates among endoscopists vary considerably, ranging from 22.4% to 52.9%. Concurrently, as the number of biopsies increases, so does the number of histologically negative biopsies from endoscopically normal-appearing mucosa, imposing an unnecessary burden on both patients and healthcare systems \u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e.Third, the accuracy of biopsy is dependent on the number and location of the samples obtained. Endoscopic biopsy is an invasive procedure that can cause mucosal damage and bleeding. Repeated biopsies or taking multiple samples over a large area may even lead to submucosal fibrosis \u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eWe aim to enhance the real-time dynamic endoscopic diagnostic accuracy for ulcer characterization and reduce unnecessary biopsies. Emerging technologies offer promising avenues to achieve this goal. Artificial intelligence (AI) has demonstrated considerable potential in auxiliary diagnosis: in recent years, AI models based on deep learning (DL) have made significant strides in medical image analysis. Current studies have indicated the applicability of AI models in automatically assessing the malignant potential of gastric ulcer images. However, most of these studies rely on single-center data and focus on classifying gastric conditions (e.g:normal mucosa, benign ulcers, malignant ulcers) based on individual static images. This approach diverges from the clinical reality, where endoscopists make diagnoses through continuous observation and holistic assessment of the entire lesion \u003csup\u003e\u003cspan additionalcitationids=\"CR15 CR16\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. Therefore, we plan to develop an AI system using multi-center data to dynamically identify gastric ulcer lesions in real-time during endoscopy. This system will perform benign versus malignant classification and provide pixel-level delineation of the lesions, thereby improving the accuracy of ulcer identification and offering guidance on the necessity of biopsy.\u003c/p\u003e"},{"header":"2 Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Study design\u003c/h2\u003e \u003cp\u003eA brief summary of the study design is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.This is a multicenter, retrospective study. All endoscopic images and videos used in this research were retrospectively collected from patients who underwent gastroscopy during routine clinical practice between January 1, 2021, and April 30, 2025, at four tertiary care hospitals in China. The participating hospitals are as follows: The Second Affiliated Hospital of Chongqing Medical University (Chongqing, China), Chongqing University Three Gorges Hospital (Chongqing, China), Chongqing Kaizhou District People's Hospital (Chongqing, China), and Chengdu Third People's Hospital (Chengdu, China). All images and videos were captured using standard endoscopes (GIF-Q260, GIF-H260, GIF-H290, Olympus Medical Systems Co. Ltd and HD-500,HD-550, SonoScape Biomedical Technology Co., Ltd). The study was approved by the respective Institutional Review Boards and was conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived by the Institutional Review Boards due to the retrospective nature of the study and the use of fully anonymized data.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Datasets and preprocessing\u003c/h2\u003e \u003cp\u003eThe inclusion criteria for images were as follows: 1) Patients aged over 18 who underwent gastroscopy and were histologically diagnosed with benign or malignant ulcerative lesions according to the World Health Organization (WHO) classification; 2) Endoscopic images without ulcerative lesions were included as the \"normal gastric mucosa\" dataset. The exclusion criteria were: 1) Cases where pathological diagnosis recommended further investigations (such as immunohistochemistry) for definitive characterization, but the patients did not complete these tests; 2) Endoscopic image quality insufficient for reliable endoscopic diagnosis; 3) Inadequate pathological diagnosis due to issues in tissue sampling, sectioning, or staining.\u003c/p\u003e \u003cp\u003eAfter selection, a total of 2,028 subjects were included. Additionally, 11 gastroscopy videos from the Second Affiliated Hospital of Chongqing Medical University were obtained, and images were extracted from the real-time video streams. This process ultimately yielded 9,840 benign ulcer images, 1,727 malignant ulcer images, and 15,791 normal background images. All images were divided into training, testing, and validation sets in an 8:1:1 ratio. These representative images were used for model development and validation.A detailed summary of the number and distribution of each category in the datasets is shown in the Supplementary table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and 2 .Three physicians, each with over five years of experience in endoscopic examination, annotated the boundaries of each lesion in the images. Two additional endoscopists with more than ten years of endoscopic experience reviewed all annotated images. Any disagreements were resolved through discussion until a consensus was reached. Pathological results were reviewed by experienced pathologists specializing in gastric abnormalities, each possessing over ten years of diagnostic experience \u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e.For image preprocessing, an object detection model was trained to automatically extract the Region of Interest (ROI) from the output images. The overall procedure is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. All final images were output in JPG format.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Building of AI Model\u003c/h2\u003e \u003cp\u003eA lightweight You Only Look Once Version 8 (YOLOv8) model was employed to construct a real-time gastric ulcer instance segmentation system, which simultaneously classifies the malignancy status of gastric ulcers and segments the ulcer regions. To address the impact of varying illumination conditions, a customized Illumination-Aware Attention module was integrated into the original YOLOv8 architecture. The model was developed using the Python programming language and the PyTorch framework, and training was conducted on three NVIDIA GeForce RTX 3080 GPUs (each with 24GB memory). More detailed information regarding the methodologies is provided in the Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e and S2.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Model evaluation and Statistical methods\u003c/h2\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.4.1 Classification performance evaluation\u003c/h2\u003e \u003cp\u003eBased on the test set, an N\u0026times;N matrix (N\u0026thinsp;=\u0026thinsp;3: normal mucosa, benign ulcer, malignant ulcer) was constructed. The rows of the matrix represent the categories marked by the pathological gold standard, and the columns represent the categories predicted by the model. The corresponding precision, specificity and recall were calculated.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.4.2 Real-time evaluation\u003c/h2\u003e \u003cp\u003eLatency per Frame: The time required to process a single image from input to output(ms).\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cp\u003e3.1 When the AI model detects gastric ulcer lesions from the input data of the test image, the AI model will output the disease name (benign ulcer or malignant ulcer) and its probability score (range 0–1). The higher the probability score, the higher the confidence in the diagnosis of the model, and the detected lesions will be outlined on the image. Meanwhile, the heat map provided by Gradient-weighted class activation mapping(Grad-CAM) indicates that our model makes reasonable and objective judgments based on specific parts of the image, similar to how endoscopists diagnose lesions. Representative cases are given in Fig.\u0026nbsp;3 and Figure S3 .\u003c/p\u003e\n\u003cp\u003e3.2 The actual number of pictures in the validation set was 2713, among which the number of lesions of\u003c/p\u003e\n\u003cp\u003enormal gastric ulcer, benign gastric ulcer and malignant gastric ulcer was 1482, 1133 and 154 respectively. Table\u0026nbsp;1 shows that the precision rate, recall and specificity of artificial intelligence in lesion recognition are 0.91, 0.91and 0.95, among which the precision rate, recall and specificity of benign lesion recognition are 0.83, 0.94 and 0.89 and the precision rate, recall and specificity of malignant lesion recognition are all 0.90, 0.91 and 0.99.The more detailed data are presented in Fig.\u0026nbsp;4.\u003c/p\u003e\n\u003cdiv\u003e\n \u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 1\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eA summary of detection performance of the AI model\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"4\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/th\u003e\n \u003cth align=\"left\"\u003ePrecision\u003cbr\u003e\u003c/th\u003e\n \u003cth align=\"left\"\u003eRecall\u003cbr\u003e\u003c/th\u003e\n \u003cth align=\"left\"\u003eSpecificity\u003cbr\u003e\u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003eBenign ulcer\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"char\"\u003e0.83\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"char\"\u003e0.94\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"char\"\u003e0.89\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003eMalignant ulcer\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"char\"\u003e0.90\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"char\"\u003e0.91\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"char\"\u003e0.99\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003eBackground\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"char\"\u003e0.96\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"char\"\u003e0.88\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"char\"\u003e0.95\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003eAll\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"char\"\u003e0.91\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"char\"\u003e0.91\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"char\"\u003e0.95\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e3.3 To evaluate the clinical real-time performance of the model, we recorded the throughput on the validation set. Table\u0026nbsp;2 shows the speed of artificial intelligence in lesion recognition and processing.\u003c/p\u003e\n\u003cdiv\u003e\n \u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 2\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eA summary of real-time performance of the AI model .\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"4\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003eNumber\u003cbr\u003e\u003c/th\u003e\n \u003cth align=\"left\"\u003eTime(s)\u003cbr\u003e\u003c/th\u003e\n \u003cth align=\"left\"\u003eSpeed(fps)\u003cbr\u003e\u003c/th\u003e\n \u003cth align=\"left\"\u003elatency(ms)\u003cbr\u003e\u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e2713\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"left\"\u003e24\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"left\"\u003e113\u003cbr\u003e\u003c/td\u003e\n \u003ctd align=\"left\"\u003e8.84\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eIn this multicenter retrospective study, we developed and validated a deep learning-based artificial intelligence model for the differential diagnosis of benign and malignant gastric ulcers. The model achieved an overall accuracy of 0.91 with a single-frame latency of only 8.84 ms.This research not only provides a novel tool for precise diagnosis of gastric ulcers but also establishes a technical pathway for optimizing clinical biopsy decisions and reducing healthcare costs. The system holds particular significance for less experienced endoscopists and challenging ulcer lesions, demonstrating potential to enhance the quality of endoscopic examinations and bridge diagnostic disparities between primary care hospitals and advanced medical centers.\u003c/p\u003e \u003cp\u003eArtificial intelligence systems have demonstrated impressive capabilities in medical image analysis. Regarding diagnostic models for gastric ulcer lesions, most previous studies were developed using single-center datasets based on individual static images. However, heterogeneity and diversity across datasets can significantly impact AI performance, making datasets that adequately represent real-world variations crucial \u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Our dataset, sourced from four different tertiary hospitals and constructed by extracting frames from real procedure videos, simulates the endoscopist's process of adjusting perspectives and observing dynamic changes in lesions during examination. This design enables the model to capture subtle features often overlooked in static images (such as border motility and adhesion stability of the exudate). For instance, the exudate covering benign ulcers may partially detach after air insufflation, whereas in malignant ulcers, it typically adheres more firmly due to infiltrative growth. Such dynamic characteristics are effectively captured through video frame extraction, reflecting the clinical reasoning process of endoscopists during gastroscopy and thereby enhancing the representativeness of our samples in real-world clinical scenarios.Furthermore, we incorporated Grad-CAM as a post-modeling technique to provide visual explanations. By deconstructing the AI model's diagnostic logic through both post-modeling and pre-modeling approaches, we achieved concretization of abstract theories and improved interpretability of the AI diagnostics, thereby enhancing diagnostic accuracy and consistency \u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eAdditionally, some images collected from older endoscopic systems exhibited low brightness. To address the common issue of uneven illumination in endoscopic images \u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e,\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e,this study introduced an illumination-aware attention module into the YOLOv8 architecture. This module dynamically integrates original features with normalized features through feature channel normalization and a differentiable gating mechanism, effectively suppressing interference from reflective areas on the model's judgment. This approach partially mitigates the selection bias often faced in retrospective studies.Ultimately, our AI model achieved an overall diagnostic accuracy of 0.91, which is comparable to the performance of expert endoscopists reported in related literature and superior to that of novice endoscopists \u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eMoreover, in the selection of optimization strategies for medical imaging AI models, the field of digestive endoscopy presents clinical requirements distinctly different from other imaging modalities. Traditional auxiliary diagnostics in imaging\u0026mdash;such as radiological image analysis and pathological slide recognition\u0026mdash;typically employ offline batch processing modes, with optimization objectives focused on maximizing system throughput to enhance overall processing efficiency. However, in real-time diagnostic and therapeutic scenarios of digestive endoscopy, due to the time-accumulated effects of anesthesia risk and the spatial constraints of endoscopic manipulation \u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e,clinical decision-making operates within strict time windows\u0026mdash;when endoscopists identify a suspicious ulcer lesion, they must complete a series of decisions including determining the nature of the lesion, assessing the necessity for biopsy, and accurately locating it within a short timeframe. This requirement for real-time interaction demands that auxiliary diagnostic models prioritize the optimization of single-frame processing latency. By reducing the inference time per frame, the cumulative effect of operational delays can be avoided, thereby ensuring the continuity of the diagnostic-therapeutic process and the timeliness of decisions.In previous studies, the total time required to complete endoscopy with or without AI assistance has been used as one of the outcome measures. These studies found no statistically significant difference in procedure time between the AI-assisted and manual groups, but this does not fully reflect the real-time performance of the model \u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. The human eye perceives a sequence of images as a continuously moving scene when the frame rate reaches 16\u0026ndash;24 frames per second (fps). Therefore, the system must possess efficient image processing capabilities, with a real-time processing efficiency of \u0026gt;\u0026thinsp;16 frames/second, and the image delay should be \u0026lt;\u0026thinsp;50 ms compared to the original endoscopic video system. This ensures that physicians can observe and diagnose seamlessly, receiving timely real-time feedback and analysis results, thus fulfilling the requirements for dynamic recognition\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. Consequently, we recorded the processing time of our AI model for each frame in the data stream, and the results demonstrate that our model exhibits excellent real-time performance.\u003c/p\u003e \u003cp\u003eThere are some limitations in this study.First, the retrospective design inherently makes it difficult to eliminate selection bias. Second, while the multi-center data source enriches sample diversity, previous research indicates that multi-center data often leads to greater diversity in negative cases but less diversity in positive cases, which may attenuate the identification of positive lesions\u0026mdash;a trend consistent with our findings \u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. In the future, we will further optimize the sample size and ratio and employ novel algorithms to overcome this issue. Additionally, all included medical institutions in this study are tertiary hospitals. We should appropriately incorporate case information from primary hospitals in resource-limited areas for model training and validation, which would be more conducive to the future validation and generalization of the AI system.\u003c/p\u003e \u003cp\u003eThird, our study is based on an end-to-end algorithm, whose diagnostic process remains an opaque and unexplainable \"black box.\" Although we adopted post-modeling and pre-modeling approaches during model development to enhance medical interpretability, the issue of explainability in AI systems remains not fully resolved, which to some extent affects the credibility and acceptability of the AI system in clinical practice. Fourth, while we treated videos as image sequences and used frame extraction to partially replicate the real examination process, it is essential to validate the model's clinical applicability for real-time diagnosis using datasets containing complete videos. We currently lack prospective experiments to further verify this aspect.\u003c/p\u003e \u003cp\u003eIn summary, this study successfully developed an artificial intelligence model capable of real-time dynamic differentiation between benign and malignant gastric ulcers by utilizing multicenter data and an improved YOLOv8 architecture. The model demonstrated excellent performance in detecting retrospective images and videos, and despite existing data and technical limitations, it has shown potential to transform traditional endoscopic diagnostic workflows. Future research should focus on multimodal integration, prospective validation, and the establishment of ethical frameworks to promote its comprehensive transition from technological exploration to routine clinical application.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eartificial intelligence\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDL\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003edeep learning\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eWHO\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eWorld Health Organization\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eROI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eRegion of Interest\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eYOLOv8\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eYou Only Look Once Version 8\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eGrad-CAM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eGradient-weighted class activation mapping\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eFPS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eframes per second\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study protocol, including the request for waiver of informed consent, was initially reviewed and approved by the\u0026nbsp;Institutional Review Board\u0026nbsp;(The Ethics Committee of the Second Affiliated Hospital of Chongqing Medical University; Approval No. 2024 Ethical Review No.129). The requirement for informed consent was waived by the Institutional Review Boards due to the retrospective nature of the study and the use of fully anonymized data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNo funding was received for this research\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors contributed to the study conception and design.Yibo Tan\u0026nbsp;：Conceptualization, Data Curation, Statistic Analysis,Writing - Original Draft；YongJun Wu\u0026nbsp;：Software，Original Draft；Junyu Lu\u0026nbsp;，Zhihang Zhou，SongHe：Conceptualization, \u0026nbsp;Writing- Reviewing and Editing and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eLanas A, Chan FKL. Peptic ulcer disease. Lancet (London, England). Aug. 2017;5(10094):613\u0026ndash;24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSmyth EC, Nilsson M, Grabsch HI, van Grieken NC, Lordick F. Gastric cancer. Lancet (London, England). Aug. 2020;29(10251):635\u0026ndash;48.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRen J, Jin X, Li J, et al. The global burden of peptic ulcer disease in 204 countries and territories from 1990 to 2019: a systematic analysis for the Global Burden of Disease Study 2019. Int J Epidemiol Oct. 2022;13(5):1666\u0026ndash;76.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKatai H, Ishikawa T, Akazawa K et al. Jan. Five-year survival analysis of surgically resected gastric cancer cases in Japan: a retrospective analysis of more than 100,000 patients from the nationwide registry of the Japanese Gastric Cancer Association (2001\u0026ndash;2007). Gastric cancer: official journal of the International Gastric Cancer Association and the Japanese Gastric Cancer Association. 2018;21(1):144\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChinese Medical Association, Chinese Medical Association Journal Press, Chinese Medical Association Digestive Disease Branch, Chinese Medical Association General Practice Branch, Editorial Committee of Chinese Journal of General Practice, Working Group for the Development of Primary Care Guidelines for Digestive System Diseases. Primary Care Guidelines for Peptic Ulcer. (2023). Chinese Journal of General Practice. 2023;22(11):1108\u0026ndash;1117.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVakil N. Peptic Ulcer Disease: A Review. Jama Dec. 2024;3(21):1832\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGielisse EA, Kuyvenhoven JP. Follow-up endoscopy for benign-appearing gastric ulcers has no additive value in detecting malignancy: It is time to individualise surveillance endoscopy. Gastric cancer: official J Int Gastric Cancer Association Japanese Gastric Cancer Association Oct. 2015;18(4):803\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDooley CP, Larson AW, Stace NH, et al. Double-contrast barium meal and upper gastrointestinal endoscopy. A comparative study. Annals of internal medicine. Oct. 1984;101(4):538\u0026ndash;45.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMenon S, Trudgill N. How commonly is upper gastrointestinal cancer missed at endoscopy? A meta-analysis. Endoscopy Int open Jun. 2014;2(2):E46\u0026ndash;50.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYalamarthi S, Witherspoon P, McCole D, Auld CD. Missed diagnoses in patients with upper gastrointestinal cancers. Endoscopy Oct. 2004;36(10):874\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJanuszewicz W, Wieszczy P, Bialek A, et al. Endoscopist biopsy rate as a quality indicator for outpatient gastroscopy: a multicenter cohort study with validation. Gastrointest endoscopy Jun. 2019;89(6):1141\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGraham DY, Schwartz JT, Cain GD, Gyorkey F. Prospective evaluation of biopsy number in the diagnosis of esophageal and gastric carcinoma. Gastroenterol Feb. 1982;82(2):228\u0026ndash;31.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBang CS, Baik GH, Kim JH, et al. Effect of Training in Upper Endoscopic Biopsy. Korean J Helicobacter Up Gastrointest Res. 2015;3(1):33\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKlang E, Barash Y, Levartovsky A, Barkin Lederer N, Lahat A. Differentiation Between Malignant and Benign Endoscopic Images of Gastric Ulcers Using Deep Learning. Clin Exp Gastroenterol. 2021;14:155\u0026ndash;62.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNam JY, Chung HJ, Choi KS, et al. Deep learning model for diagnosing gastric mucosal lesions using endoscopic images: development, validation, and method comparison. Gastrointest endoscopy Feb. 2022;95(2):258\u0026ndash;e268210.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNamikawa K, Hirasawa T, Nakano K, et al. Artificial intelligence-based diagnostic system classifying gastric cancers and ulcers: comparison between the original and newly developed systems. Endoscopy Dec. 2020;52(12):1077\u0026ndash;83.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Y, Zhang L, Hao Z, et al. An xception model based on residual attention mechanism for the classification of benign and malignant gastric ulcers. Sci Rep Sep. 2022;13(1):15365.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiZhaoshen. Experts consensus on quality control system of data collection and labeling for artificial intelligence in digestive endoscopy (2019, Shanghai). Chin J Dig Endoscopy.37(8):533\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNorgeot B, Quer G, Beaulieu-Jones BK, et al. Minimum information about clinical artificial intelligence modeling: the MI-CLAIM checklist. Nat Med Sep. 2020;26(9):1320\u0026ndash;4.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRyan M, Stahl BC. Artificial intelligence ethics guidelines for developers and users: clarifying their content and normative implications. J Inform Communication Ethics Soc. 2020;19(1):61\u0026ndash;86.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDong Z, Wang J, Li Y, et al. Explainable artificial intelligence incorporated with domain knowledge diagnosing early gastric neoplasms under white light endoscopy. NPJ digital medicine. Apr. 2023;12(1):64.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBi C, Liu B, Wang T et al. Gastric Mucosal Blood Flow Analysis of Ulcer-healing by Electronic Endoscope System.Chin. J Gastroenterol. 2000(04):240\u0026ndash;2.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eExtracting clinical information from endoscopic capsule exams using MPEG-7 visual descriptors. 2nd European Workshop on the Integration of Knowledge, Semantics and Digital Media Technology (EWIMT 2005):105\u0026ndash;110.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYuan XL, Zhou Y, Liu W, et al. Artificial intelligence for diagnosing gastric lesions under white-light endoscopy. Surg endoscopy Dec. 2022;36(12):9444\u0026ndash;53.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eManno M, Deiana S, Gabbani T, et al. Implementation of the European Society of Gastrointestinal Endoscopy (ESGE) and European Society of Gastroenterology and Endoscopy Nurses and Associates (ESGENA) sedation training course in a regular endoscopy unit. Endoscopy Jan. 2021;53(1):65\u0026ndash;71.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu C, Zhu Y, Wu L, et al. Evaluating the effect of an artificial intelligence system on the anesthesia quality control during gastrointestinal endoscopy with sedation: a randomized controlled trial. BMC anesthesiology Oct. 2022;7(1):313.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYuan XL, Liu W, Lin YX, et al. Effect of an artificial intelligence-assisted system on endoscopic diagnosis of superficial oesophageal squamous cell carcinoma and precancerous lesions: a multicentre, tandem, double-blind, randomised controlled trial. The lancet. Gastroenterol Hepatol Jan. 2024;9(1):34\u0026ndash;44.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBig Data Collaboration Group, Digestive Endoscopology Branch of Chinese Medical Association. Expert consensus on the clinical application of artificial intelligence system to upper gastrointestinal endoscopy (2023, Wuhan). Chin J Dig Endoscopy. 2024;41(2):85\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDong Z, Tao X, Du H, et al. Exploring the challenge of early gastric cancer diagnostic AI system face in multiple centers and its potential solutions. J Gastroenterol Oct. 2023;58(10):978\u0026ndash;89.\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":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-gastroenterology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmge","sideBox":"Learn more about [BMC Gastroenterology](http://bmcgastroenterol.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmge/default.aspx","title":"BMC Gastroenterology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Artificial Intelligence, Gastric Ulcer, Benign and Malignant Differentiation, Deep Learning, Endoscopic Diagnosis","lastPublishedDoi":"10.21203/rs.3.rs-8776954/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8776954/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eAim: \u003c/strong\u003eTo develop and validate a deep learning-based AI system for the dynamic, real-time differentiation of benign and malignant gastric ulcers during endoscopy, with the goal of enhancing diagnostic precision and reducing superfluous biopsies.\u003c/p\u003e\n\u003cp\u003eMethods: This was a multicenter, retrospective study collecting endoscopic images and videos from four tertiary hospitals in China. An improved YOLOv8 model, incorporating an illumination attention module, was developed for real-time instance segmentation and classification. The dataset comprised 9,840 benign ulcer images, 1,727 malignant ulcer images, and 15,791 normal mucosa images, split into training, testing, and validation sets at an 8:1:1 ratio. Performance was evaluated based on precision, recall, specificity, and processing latency.\u003c/p\u003e\n\u003cp\u003eResults: On the validation set, the AI model achieved an overall precision, recall, and specificity of 0.91, 0.91, and 0.95, respectively. For malignant ulcer recognition specifically, the precision, recall, and specificity were 0.90, 0.91, and 0.99. The model demonstrated strong real-time performance with a latency of 8.84 ms per frame and a processing speed of 113 frames per second.\u003c/p\u003e\n\u003cp\u003eConclusion: The developed AI model enables accurate, real-time discrimination between benign and malignant gastric ulcers during endoscopy. It holds significant potential to augment clinical decision-making, standardize diagnostic quality, and optimize biopsy strategies.\u003c/p\u003e","manuscriptTitle":"Development and validation of a real-time AI model for differentiating benign and malignant gastric ulcers: A multicenter retrospective study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-17 11:45:37","doi":"10.21203/rs.3.rs-8776954/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-03-10T05:24:01+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-05T13:58:42+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-03T08:16:10+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"257783531503310691447207898337873157087","date":"2026-02-28T04:05:10+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"60478527823374493964309567674561074139","date":"2026-02-28T02:28:33+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-11T17:18:05+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-06T10:17:50+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-05T13:19:32+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Gastroenterology","date":"2026-02-05T12:49:43+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-gastroenterology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmge","sideBox":"Learn more about [BMC Gastroenterology](http://bmcgastroenterol.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmge/default.aspx","title":"BMC Gastroenterology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8163827e-0d44-40b8-8e7b-079674636631","owner":[],"postedDate":"February 17th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-15T09:50:36+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-17 11:45:37","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8776954","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8776954","identity":"rs-8776954","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}