User login
Dear colleagues,
Since our last Perspectives feature on artificial intelligence (AI) in gastroenterology and hepatology, the field has experienced remarkable growth in both innovation and clinical adoption. AI tools that were once conceptual are now entering everyday practice, with many more on the horizon poised to transform how we diagnose, treat, and manage patients.
Dr. Yuvaraj Singh, Dr. Alessandro Colletta, and Dr. Neil Marya discuss how purpose-built AI models can reduce diagnostic uncertainty in advanced endoscopy. From cholangioscopy systems that outperform standard ERCP sampling in distinguishing malignant biliary strictures to EUS-based platforms that differentiate autoimmune pancreatitis from pancreatic cancer, they envision a near-term future in which machine intelligence enhances accuracy, accelerates decision-making, and refines interpretation—without replacing the clinician’s expertise.
Complementing this, Dr. Dennis Shung takes a broader view across the endoscopy unit and outpatient clinic. He highlights the promise of AI for polyp detection, digital biopsy, and automated reporting, while underscoring the importance of human oversight, workflow integration, and safeguards against misinformation. Dr. Shung also emphasizes the pivotal role professional societies can play in establishing clear standards, ethical boundaries, and trusted frameworks for AI deployment in GI practice.
We hope these perspectives spark practical conversations about when—and how—to integrate AI in your own practice. As always, we welcome your feedback and real-world experience. Join the conversation on X at @AGA_GIHN.
Gyanprakash A. Ketwaroo, MD, MSc, is associate professor of medicine, Yale University, New Haven, and chief of endoscopy at West Haven VA Medical Center, both in Connecticut. He is an associate editor for GI & Hepatology News.
AI Models in Advanced Endoscopy
BY YUVARAJ SINGH, MD; ALESSANDRO COLLETTA, MD; NEIL MARYA, MD
As the adage goes, “if tumor is the rumor, then tissue is the issue, because cancer may be the answer.”
Establishing an accurate diagnosis is the essential first step toward curing or palliating malignancy. From detecting an early neoplastic lesion, to distinguishing between malignant and benign pathology, or to determining when and where to obtain tissue, endoscopists are frequently faced with the challenge of transforming diagnostic suspicion into certainty.
Artificial intelligence (AI), designed to replicate human cognition such as pattern recognition and decision-making, has emerged as a technology to assist gastroenterologists in addressing a variety of different tasks during endoscopy. AI research in gastrointestinal endoscopy has initially focused on computer-aided detection (CADe) of colorectal polyps. More recently, however, there has been increased emphasis on developing AI to assist advanced endoscopists.
For instance, in biliary endoscopy, AI is being explored to improve the notoriously challenging diagnosis of cholangiocarcinoma, where conventional tissue sampling often falls short of providing a definitive diagnosis. Similarly, in the pancreas, AI models are showing potential to differentiate autoimmune pancreatitis (AIP) from pancreatic ductal adenocarcinoma (PDAC), a distinction with profound therapeutic implications. Even pancreatic cysts are beginning to benefit from AI models that refine risk stratification and guide management. Together, these advances underscore how AI is not merely an adjunct but a potentially massive catalyst for reimagining the diagnostic role of advanced endoscopists.
Classifying biliary strictures (MBS) accurately remains a challenge. Standard ERCP-based sampling techniques (forceps biopsy and brush cytology) are suboptimal diagnostic tools with false negative rates for detecting MBS of less than 50%. The diagnostic uncertainty related to biliary stricture classification carries significant consequences for patients. For example, patients with biliary cancer without positive cytology have treatments delayed until a malignant diagnosis is established.
Ancillary technologies to enhance ERCP-based tissue acquisition are still weighed down by low sensitivity and accuracy; even with ancillary use of fluorescent in situ hybridization (FISH), diagnostic yield remains limited. EUS-FNA can help with distal biliary strictures, but this technique risks needle-tract seeding in cases of perihilar disease. Cholangioscopy allows for direct visualization and targeted sampling; however, cholangioscopy-guided forceps biopsies are burdened by low sensitivities.1 Additionally, physician interpretation of visual findings during cholangioscopy often suffers from poor interobserver agreement and poor accuracy.2
To improve the classification of biliary strictures, several groups have studied the application of AI for cholangioscopy footage of biliary pathology. In our lab, we trained an AI incorporating over 2.3 million cholangioscopy still images and nearly 20,000 expert-annotated frames to enhance its development. The AI closely mirrored expert labeling of cholangioscopy images of malignant pathology and, when tested on full cholangioscopy videos of indeterminate biliary strictures, the AI achieved a diagnostic accuracy of 91%—outperforming both brush cytology (63%) and forceps biopsy (61%).3
The results from this initial study were later validated across multiple centers. AI-assisted cholangioscopy could thus offer a reproducible, real-world solution to one of the most persistent diagnostic dilemmas advanced endoscopists face—helping clinicians act earlier and with greater confidence when evaluating indeterminate strictures.
Moving from the biliary tree to the pancreas, autoimmune pancreatitis (AIP) is a benign fibro-inflammatory disease that often frustrates advanced endoscopists as it closely mimics the appearance of pancreatic ductal adenocarcinoma (PDAC). The stakes are high: despite modern diagnostic techniques, including advanced imaging, some patients with pancreatic resections for “suspected PDAC” are still found to have AIP on final pathology. Conventional tools to distinguish AIP from PDAC have gaps: serum IgG4 and EUS-guided biopsies are both specific but insensitive.
Using EUS videos and images of various pancreas pathologies at Mayo Clinic, we developed an AI to tackle this dilemma. After intensive training, the EUS AI achieved a greater accuracy for distinguishing AIP from PDAC than a group of expert Mayo clinic endosonographers.5 In practice, an EUS-AI can identify AIP patterns in real-time, guiding clinicians toward steroid trials or biopsies and reducing the need for unnecessary surgeries.
Looking ahead, there are multiple opportunities for integration of AI into advanced endoscopy practices. Ongoing research suggests that AI could soon assist with identification of pancreas cysts most at risk for malignant transformation, classification of high risk Barrett’s esophagus, and even help with rapid on-site assessment of cytologic specimens obtained during EUS. Beyond diagnosis, AI could likely play an important role in guiding therapeutic interventions. For example, an ERCP AI in the future may be able to provide cannulation assistance or an AI assistant could help endosonographers during deployments of lumen apposing metal stents.
By enhancing image interpretation and procedural consistency, AI has the potential to uphold the fundamental principle of primum non nocere, enabling us to intervene with precision while minimizing harm. AI can also bridge grey zones in clinical practice and narrow diagnostic uncertainty in real time. Importantly, these systems can help clinicians achieve expertise in a fraction of the time it traditionally takes to acquire comparable human proficiency, while offering wider availability across practice settings and reducing interobserver variability that has long challenged endoscopic interpretation.
Currently, adoption is limited by high bias risk, lack of external validation, and interpretability Still, the trajectory of AI suggests a future where these computer technologies will not only support but also elevate human expertise, reshaping the standards of care of diseases managed by advanced endoscopists.
Dr. Singh, Dr. Colletta, and Dr. Marya are based at the Division of Gastroenterology and Hepatology, UMass Chan Medical School, Worcester, Massachusetts. Dr. Marya is a consultant for Boston Scientific, and has no other disclosures. Dr. Singh and Dr. Colletta have no disclosures.
References
1. Navaneethan U, et al. Comparative effectiveness of biliary brush cytology and intraductal biopsy for detection of malignant biliary strictures: a systematic review and meta-analysis. Gastrointest Endosc. 2015 Jan. doi: 10.1016/j.gie.2014.09.017.
2. Stassen PMC, et al. Diagnostic accuracy and interobserver agreement of digital single-operator cholangioscopy for indeterminate biliary strictures. Gastrointest Endosc 2021 Dec. doi: 10.1016/j.gie.2021.06.027.
3. Marya NB, et al. Identification of patients with malignant biliary strictures using a cholangioscopy-based deep learning artificial intelligence (with video). Gastrointest Endosc. 2023 Feb. doi: 10.1016/j.gie.2022.08.021.
4. Marya NB, et al. Multicenter validation of a cholangioscopy artificial intelligence system for the evaluation of biliary tract disease. Endoscopy. 2025 Aug. doi: 10.1055/a-2650-0789.
5. Marya NB, et al. Utilisation of artificial intelligence for the development of an EUS-convolutional neural network model trained to enhance the diagnosis of autoimmune pancreatitis. Gut. 2021 Jul. doi: 10.1136/gutjnl-2020-322821.
AI in General GI and Endoscopy
BY DENNIS L. SHUNG, MD, MHS, PHD
The practice of gastroenterology is changing, but much of it will be rooted in the same – careful, focused attention on endoscopic procedures, and compassionate, attentive care in clinic. Artificial intelligence (AI), like the Industrial Revolution before, is going to transform our practice. This comes with upsides and downsides, and highlights the need for strong leadership from our societies to safeguard the technology for practitioners and patients.
What are the upsides?
AI has the potential to serve as a second set of eyes in detecting colon polyps, increasing the adenoma detection rate (ADR).1 AI can be applied to all areas of the gastrointestinal tract, providing digital biopsies, guiding resection, and ensuring quality, which are all now possible with powerful new endoscopy foundation models, such as GastroNet-5M.2
Additionally. the advent of automating the collection of data into reports may herald the end of our days as data entry clerks. Generative AI also has the potential to give us all the best information at our fingertips, suggesting guideline-based care, providing the most up to date evidence, and guiding the differential diagnosis. The potential for patient-facing AI systems could lead to better health literacy, more meaningful engagement, and improved patient satisfaction.3
What are the downsides?
For endoscopy, AI cannot make up for poor technique to ensure adequate mucosal exposure by the endoscopist, and an increase in AI-supported ADR does not yet convincingly translate into concrete gains in colorectal cancer-related mortality. For the foreseeable future, AI cannot make a connection with the patient in front of us, which is critical in diagnosing and treating patients.
Currently, AI appears to worsen loneliness4, and does not necessarily deepen the bonds or provide the positive touch that can heal, and which for many of us, was the reason we became physicians. Finally, as information proliferates, the information risk to patients and providers is growing – in the future, trusted sources to monitor, curate, and guide AI will be ever more important.
Black Swans
As AI begins to mature, there are risks that lurk beneath the surface. When regulatory bodies begin to look at AI-assisted diagnostics or therapeutics as the new standard of care, reimbursement models may adjust, and providers may be left behind. The rapid proliferation and haphazard adoption of AI could lead to overdependence and deskilling or result in weird and as yet unknown errors that are difficult to troubleshoot.
What is the role of the GI societies?
Specialty societies like AGA are taking leadership roles in determining the bounds of where AIs may tread, not just in providing information to their membership but also in digesting evidence and synthesizing recommendations. Societies must balance the real promise of AI in endoscopy with the practice realities for members, and provide living guidelines that reflect the consensus of members regarding scope of practice with the ability to update as new data become available.5
Societies also have a role as advocates for safety, taking ownership of high-quality content to prevent misinformation. AGA recently announced the development of a chat interface that will be focused on providing its members the highest quality information, and serve as a portal to identify and respond to its members’ information needs. By staying united rather than fragmenting, societies can maintain bounds to protect its members and their patients and advance areas where there is clinical need, together.
Dr. Shung is senior associate consultant, Division of Gastroenterology and Hepatology, and director of clinical generative artificial intelligence and informatics, Department of Medicine, at Mayo Clinic Rochester, Minnesota. He has no disclosures in regard to this article.
References
1. Soleymanjahi S, et al. Artificial Intelligence-Assisted Colonoscopy for Polyp Detection : A Systematic Review and Meta-analysis. Ann Intern Med. 2024 Dec. doi:10.7326/annals-24-00981.
2. Jong MR, et al. GastroNet-5M: A Multicenter Dataset for Developing Foundation Models in Gastrointestinal Endoscopy. Gastroenterology. 2025 Jul. doi: 10.1053/j.gastro.2025.07.030.
3. Soroush A, et al. Generative Artificial Intelligence in Clinical Medicine and Impact on Gastroenterology. Gastroenterology. 2025 Aug. doi: 10.1053/j.gastro.2025.03.038.
4. Mengying Fang C, et al. How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study. arXiv e-prints. 2025 Mar. doi: 10.48550/arXiv.2503.17473.
5. Sultan S, et al. AGA Living Clinical Practice Guideline on Computer-Aided Detection-Assisted Colonoscopy. Gastroenterology. 2025 Apr. doi:10.1053/j.gastro.2025.01.002.
Dear colleagues,
Since our last Perspectives feature on artificial intelligence (AI) in gastroenterology and hepatology, the field has experienced remarkable growth in both innovation and clinical adoption. AI tools that were once conceptual are now entering everyday practice, with many more on the horizon poised to transform how we diagnose, treat, and manage patients.
Dr. Yuvaraj Singh, Dr. Alessandro Colletta, and Dr. Neil Marya discuss how purpose-built AI models can reduce diagnostic uncertainty in advanced endoscopy. From cholangioscopy systems that outperform standard ERCP sampling in distinguishing malignant biliary strictures to EUS-based platforms that differentiate autoimmune pancreatitis from pancreatic cancer, they envision a near-term future in which machine intelligence enhances accuracy, accelerates decision-making, and refines interpretation—without replacing the clinician’s expertise.
Complementing this, Dr. Dennis Shung takes a broader view across the endoscopy unit and outpatient clinic. He highlights the promise of AI for polyp detection, digital biopsy, and automated reporting, while underscoring the importance of human oversight, workflow integration, and safeguards against misinformation. Dr. Shung also emphasizes the pivotal role professional societies can play in establishing clear standards, ethical boundaries, and trusted frameworks for AI deployment in GI practice.
We hope these perspectives spark practical conversations about when—and how—to integrate AI in your own practice. As always, we welcome your feedback and real-world experience. Join the conversation on X at @AGA_GIHN.
Gyanprakash A. Ketwaroo, MD, MSc, is associate professor of medicine, Yale University, New Haven, and chief of endoscopy at West Haven VA Medical Center, both in Connecticut. He is an associate editor for GI & Hepatology News.
AI Models in Advanced Endoscopy
BY YUVARAJ SINGH, MD; ALESSANDRO COLLETTA, MD; NEIL MARYA, MD
As the adage goes, “if tumor is the rumor, then tissue is the issue, because cancer may be the answer.”
Establishing an accurate diagnosis is the essential first step toward curing or palliating malignancy. From detecting an early neoplastic lesion, to distinguishing between malignant and benign pathology, or to determining when and where to obtain tissue, endoscopists are frequently faced with the challenge of transforming diagnostic suspicion into certainty.
Artificial intelligence (AI), designed to replicate human cognition such as pattern recognition and decision-making, has emerged as a technology to assist gastroenterologists in addressing a variety of different tasks during endoscopy. AI research in gastrointestinal endoscopy has initially focused on computer-aided detection (CADe) of colorectal polyps. More recently, however, there has been increased emphasis on developing AI to assist advanced endoscopists.
For instance, in biliary endoscopy, AI is being explored to improve the notoriously challenging diagnosis of cholangiocarcinoma, where conventional tissue sampling often falls short of providing a definitive diagnosis. Similarly, in the pancreas, AI models are showing potential to differentiate autoimmune pancreatitis (AIP) from pancreatic ductal adenocarcinoma (PDAC), a distinction with profound therapeutic implications. Even pancreatic cysts are beginning to benefit from AI models that refine risk stratification and guide management. Together, these advances underscore how AI is not merely an adjunct but a potentially massive catalyst for reimagining the diagnostic role of advanced endoscopists.
Classifying biliary strictures (MBS) accurately remains a challenge. Standard ERCP-based sampling techniques (forceps biopsy and brush cytology) are suboptimal diagnostic tools with false negative rates for detecting MBS of less than 50%. The diagnostic uncertainty related to biliary stricture classification carries significant consequences for patients. For example, patients with biliary cancer without positive cytology have treatments delayed until a malignant diagnosis is established.
Ancillary technologies to enhance ERCP-based tissue acquisition are still weighed down by low sensitivity and accuracy; even with ancillary use of fluorescent in situ hybridization (FISH), diagnostic yield remains limited. EUS-FNA can help with distal biliary strictures, but this technique risks needle-tract seeding in cases of perihilar disease. Cholangioscopy allows for direct visualization and targeted sampling; however, cholangioscopy-guided forceps biopsies are burdened by low sensitivities.1 Additionally, physician interpretation of visual findings during cholangioscopy often suffers from poor interobserver agreement and poor accuracy.2
To improve the classification of biliary strictures, several groups have studied the application of AI for cholangioscopy footage of biliary pathology. In our lab, we trained an AI incorporating over 2.3 million cholangioscopy still images and nearly 20,000 expert-annotated frames to enhance its development. The AI closely mirrored expert labeling of cholangioscopy images of malignant pathology and, when tested on full cholangioscopy videos of indeterminate biliary strictures, the AI achieved a diagnostic accuracy of 91%—outperforming both brush cytology (63%) and forceps biopsy (61%).3
The results from this initial study were later validated across multiple centers. AI-assisted cholangioscopy could thus offer a reproducible, real-world solution to one of the most persistent diagnostic dilemmas advanced endoscopists face—helping clinicians act earlier and with greater confidence when evaluating indeterminate strictures.
Moving from the biliary tree to the pancreas, autoimmune pancreatitis (AIP) is a benign fibro-inflammatory disease that often frustrates advanced endoscopists as it closely mimics the appearance of pancreatic ductal adenocarcinoma (PDAC). The stakes are high: despite modern diagnostic techniques, including advanced imaging, some patients with pancreatic resections for “suspected PDAC” are still found to have AIP on final pathology. Conventional tools to distinguish AIP from PDAC have gaps: serum IgG4 and EUS-guided biopsies are both specific but insensitive.
Using EUS videos and images of various pancreas pathologies at Mayo Clinic, we developed an AI to tackle this dilemma. After intensive training, the EUS AI achieved a greater accuracy for distinguishing AIP from PDAC than a group of expert Mayo clinic endosonographers.5 In practice, an EUS-AI can identify AIP patterns in real-time, guiding clinicians toward steroid trials or biopsies and reducing the need for unnecessary surgeries.
Looking ahead, there are multiple opportunities for integration of AI into advanced endoscopy practices. Ongoing research suggests that AI could soon assist with identification of pancreas cysts most at risk for malignant transformation, classification of high risk Barrett’s esophagus, and even help with rapid on-site assessment of cytologic specimens obtained during EUS. Beyond diagnosis, AI could likely play an important role in guiding therapeutic interventions. For example, an ERCP AI in the future may be able to provide cannulation assistance or an AI assistant could help endosonographers during deployments of lumen apposing metal stents.
By enhancing image interpretation and procedural consistency, AI has the potential to uphold the fundamental principle of primum non nocere, enabling us to intervene with precision while minimizing harm. AI can also bridge grey zones in clinical practice and narrow diagnostic uncertainty in real time. Importantly, these systems can help clinicians achieve expertise in a fraction of the time it traditionally takes to acquire comparable human proficiency, while offering wider availability across practice settings and reducing interobserver variability that has long challenged endoscopic interpretation.
Currently, adoption is limited by high bias risk, lack of external validation, and interpretability Still, the trajectory of AI suggests a future where these computer technologies will not only support but also elevate human expertise, reshaping the standards of care of diseases managed by advanced endoscopists.
Dr. Singh, Dr. Colletta, and Dr. Marya are based at the Division of Gastroenterology and Hepatology, UMass Chan Medical School, Worcester, Massachusetts. Dr. Marya is a consultant for Boston Scientific, and has no other disclosures. Dr. Singh and Dr. Colletta have no disclosures.
References
1. Navaneethan U, et al. Comparative effectiveness of biliary brush cytology and intraductal biopsy for detection of malignant biliary strictures: a systematic review and meta-analysis. Gastrointest Endosc. 2015 Jan. doi: 10.1016/j.gie.2014.09.017.
2. Stassen PMC, et al. Diagnostic accuracy and interobserver agreement of digital single-operator cholangioscopy for indeterminate biliary strictures. Gastrointest Endosc 2021 Dec. doi: 10.1016/j.gie.2021.06.027.
3. Marya NB, et al. Identification of patients with malignant biliary strictures using a cholangioscopy-based deep learning artificial intelligence (with video). Gastrointest Endosc. 2023 Feb. doi: 10.1016/j.gie.2022.08.021.
4. Marya NB, et al. Multicenter validation of a cholangioscopy artificial intelligence system for the evaluation of biliary tract disease. Endoscopy. 2025 Aug. doi: 10.1055/a-2650-0789.
5. Marya NB, et al. Utilisation of artificial intelligence for the development of an EUS-convolutional neural network model trained to enhance the diagnosis of autoimmune pancreatitis. Gut. 2021 Jul. doi: 10.1136/gutjnl-2020-322821.
AI in General GI and Endoscopy
BY DENNIS L. SHUNG, MD, MHS, PHD
The practice of gastroenterology is changing, but much of it will be rooted in the same – careful, focused attention on endoscopic procedures, and compassionate, attentive care in clinic. Artificial intelligence (AI), like the Industrial Revolution before, is going to transform our practice. This comes with upsides and downsides, and highlights the need for strong leadership from our societies to safeguard the technology for practitioners and patients.
What are the upsides?
AI has the potential to serve as a second set of eyes in detecting colon polyps, increasing the adenoma detection rate (ADR).1 AI can be applied to all areas of the gastrointestinal tract, providing digital biopsies, guiding resection, and ensuring quality, which are all now possible with powerful new endoscopy foundation models, such as GastroNet-5M.2
Additionally. the advent of automating the collection of data into reports may herald the end of our days as data entry clerks. Generative AI also has the potential to give us all the best information at our fingertips, suggesting guideline-based care, providing the most up to date evidence, and guiding the differential diagnosis. The potential for patient-facing AI systems could lead to better health literacy, more meaningful engagement, and improved patient satisfaction.3
What are the downsides?
For endoscopy, AI cannot make up for poor technique to ensure adequate mucosal exposure by the endoscopist, and an increase in AI-supported ADR does not yet convincingly translate into concrete gains in colorectal cancer-related mortality. For the foreseeable future, AI cannot make a connection with the patient in front of us, which is critical in diagnosing and treating patients.
Currently, AI appears to worsen loneliness4, and does not necessarily deepen the bonds or provide the positive touch that can heal, and which for many of us, was the reason we became physicians. Finally, as information proliferates, the information risk to patients and providers is growing – in the future, trusted sources to monitor, curate, and guide AI will be ever more important.
Black Swans
As AI begins to mature, there are risks that lurk beneath the surface. When regulatory bodies begin to look at AI-assisted diagnostics or therapeutics as the new standard of care, reimbursement models may adjust, and providers may be left behind. The rapid proliferation and haphazard adoption of AI could lead to overdependence and deskilling or result in weird and as yet unknown errors that are difficult to troubleshoot.
What is the role of the GI societies?
Specialty societies like AGA are taking leadership roles in determining the bounds of where AIs may tread, not just in providing information to their membership but also in digesting evidence and synthesizing recommendations. Societies must balance the real promise of AI in endoscopy with the practice realities for members, and provide living guidelines that reflect the consensus of members regarding scope of practice with the ability to update as new data become available.5
Societies also have a role as advocates for safety, taking ownership of high-quality content to prevent misinformation. AGA recently announced the development of a chat interface that will be focused on providing its members the highest quality information, and serve as a portal to identify and respond to its members’ information needs. By staying united rather than fragmenting, societies can maintain bounds to protect its members and their patients and advance areas where there is clinical need, together.
Dr. Shung is senior associate consultant, Division of Gastroenterology and Hepatology, and director of clinical generative artificial intelligence and informatics, Department of Medicine, at Mayo Clinic Rochester, Minnesota. He has no disclosures in regard to this article.
References
1. Soleymanjahi S, et al. Artificial Intelligence-Assisted Colonoscopy for Polyp Detection : A Systematic Review and Meta-analysis. Ann Intern Med. 2024 Dec. doi:10.7326/annals-24-00981.
2. Jong MR, et al. GastroNet-5M: A Multicenter Dataset for Developing Foundation Models in Gastrointestinal Endoscopy. Gastroenterology. 2025 Jul. doi: 10.1053/j.gastro.2025.07.030.
3. Soroush A, et al. Generative Artificial Intelligence in Clinical Medicine and Impact on Gastroenterology. Gastroenterology. 2025 Aug. doi: 10.1053/j.gastro.2025.03.038.
4. Mengying Fang C, et al. How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study. arXiv e-prints. 2025 Mar. doi: 10.48550/arXiv.2503.17473.
5. Sultan S, et al. AGA Living Clinical Practice Guideline on Computer-Aided Detection-Assisted Colonoscopy. Gastroenterology. 2025 Apr. doi:10.1053/j.gastro.2025.01.002.
Dear colleagues,
Since our last Perspectives feature on artificial intelligence (AI) in gastroenterology and hepatology, the field has experienced remarkable growth in both innovation and clinical adoption. AI tools that were once conceptual are now entering everyday practice, with many more on the horizon poised to transform how we diagnose, treat, and manage patients.
Dr. Yuvaraj Singh, Dr. Alessandro Colletta, and Dr. Neil Marya discuss how purpose-built AI models can reduce diagnostic uncertainty in advanced endoscopy. From cholangioscopy systems that outperform standard ERCP sampling in distinguishing malignant biliary strictures to EUS-based platforms that differentiate autoimmune pancreatitis from pancreatic cancer, they envision a near-term future in which machine intelligence enhances accuracy, accelerates decision-making, and refines interpretation—without replacing the clinician’s expertise.
Complementing this, Dr. Dennis Shung takes a broader view across the endoscopy unit and outpatient clinic. He highlights the promise of AI for polyp detection, digital biopsy, and automated reporting, while underscoring the importance of human oversight, workflow integration, and safeguards against misinformation. Dr. Shung also emphasizes the pivotal role professional societies can play in establishing clear standards, ethical boundaries, and trusted frameworks for AI deployment in GI practice.
We hope these perspectives spark practical conversations about when—and how—to integrate AI in your own practice. As always, we welcome your feedback and real-world experience. Join the conversation on X at @AGA_GIHN.
Gyanprakash A. Ketwaroo, MD, MSc, is associate professor of medicine, Yale University, New Haven, and chief of endoscopy at West Haven VA Medical Center, both in Connecticut. He is an associate editor for GI & Hepatology News.
AI Models in Advanced Endoscopy
BY YUVARAJ SINGH, MD; ALESSANDRO COLLETTA, MD; NEIL MARYA, MD
As the adage goes, “if tumor is the rumor, then tissue is the issue, because cancer may be the answer.”
Establishing an accurate diagnosis is the essential first step toward curing or palliating malignancy. From detecting an early neoplastic lesion, to distinguishing between malignant and benign pathology, or to determining when and where to obtain tissue, endoscopists are frequently faced with the challenge of transforming diagnostic suspicion into certainty.
Artificial intelligence (AI), designed to replicate human cognition such as pattern recognition and decision-making, has emerged as a technology to assist gastroenterologists in addressing a variety of different tasks during endoscopy. AI research in gastrointestinal endoscopy has initially focused on computer-aided detection (CADe) of colorectal polyps. More recently, however, there has been increased emphasis on developing AI to assist advanced endoscopists.
For instance, in biliary endoscopy, AI is being explored to improve the notoriously challenging diagnosis of cholangiocarcinoma, where conventional tissue sampling often falls short of providing a definitive diagnosis. Similarly, in the pancreas, AI models are showing potential to differentiate autoimmune pancreatitis (AIP) from pancreatic ductal adenocarcinoma (PDAC), a distinction with profound therapeutic implications. Even pancreatic cysts are beginning to benefit from AI models that refine risk stratification and guide management. Together, these advances underscore how AI is not merely an adjunct but a potentially massive catalyst for reimagining the diagnostic role of advanced endoscopists.
Classifying biliary strictures (MBS) accurately remains a challenge. Standard ERCP-based sampling techniques (forceps biopsy and brush cytology) are suboptimal diagnostic tools with false negative rates for detecting MBS of less than 50%. The diagnostic uncertainty related to biliary stricture classification carries significant consequences for patients. For example, patients with biliary cancer without positive cytology have treatments delayed until a malignant diagnosis is established.
Ancillary technologies to enhance ERCP-based tissue acquisition are still weighed down by low sensitivity and accuracy; even with ancillary use of fluorescent in situ hybridization (FISH), diagnostic yield remains limited. EUS-FNA can help with distal biliary strictures, but this technique risks needle-tract seeding in cases of perihilar disease. Cholangioscopy allows for direct visualization and targeted sampling; however, cholangioscopy-guided forceps biopsies are burdened by low sensitivities.1 Additionally, physician interpretation of visual findings during cholangioscopy often suffers from poor interobserver agreement and poor accuracy.2
To improve the classification of biliary strictures, several groups have studied the application of AI for cholangioscopy footage of biliary pathology. In our lab, we trained an AI incorporating over 2.3 million cholangioscopy still images and nearly 20,000 expert-annotated frames to enhance its development. The AI closely mirrored expert labeling of cholangioscopy images of malignant pathology and, when tested on full cholangioscopy videos of indeterminate biliary strictures, the AI achieved a diagnostic accuracy of 91%—outperforming both brush cytology (63%) and forceps biopsy (61%).3
The results from this initial study were later validated across multiple centers. AI-assisted cholangioscopy could thus offer a reproducible, real-world solution to one of the most persistent diagnostic dilemmas advanced endoscopists face—helping clinicians act earlier and with greater confidence when evaluating indeterminate strictures.
Moving from the biliary tree to the pancreas, autoimmune pancreatitis (AIP) is a benign fibro-inflammatory disease that often frustrates advanced endoscopists as it closely mimics the appearance of pancreatic ductal adenocarcinoma (PDAC). The stakes are high: despite modern diagnostic techniques, including advanced imaging, some patients with pancreatic resections for “suspected PDAC” are still found to have AIP on final pathology. Conventional tools to distinguish AIP from PDAC have gaps: serum IgG4 and EUS-guided biopsies are both specific but insensitive.
Using EUS videos and images of various pancreas pathologies at Mayo Clinic, we developed an AI to tackle this dilemma. After intensive training, the EUS AI achieved a greater accuracy for distinguishing AIP from PDAC than a group of expert Mayo clinic endosonographers.5 In practice, an EUS-AI can identify AIP patterns in real-time, guiding clinicians toward steroid trials or biopsies and reducing the need for unnecessary surgeries.
Looking ahead, there are multiple opportunities for integration of AI into advanced endoscopy practices. Ongoing research suggests that AI could soon assist with identification of pancreas cysts most at risk for malignant transformation, classification of high risk Barrett’s esophagus, and even help with rapid on-site assessment of cytologic specimens obtained during EUS. Beyond diagnosis, AI could likely play an important role in guiding therapeutic interventions. For example, an ERCP AI in the future may be able to provide cannulation assistance or an AI assistant could help endosonographers during deployments of lumen apposing metal stents.
By enhancing image interpretation and procedural consistency, AI has the potential to uphold the fundamental principle of primum non nocere, enabling us to intervene with precision while minimizing harm. AI can also bridge grey zones in clinical practice and narrow diagnostic uncertainty in real time. Importantly, these systems can help clinicians achieve expertise in a fraction of the time it traditionally takes to acquire comparable human proficiency, while offering wider availability across practice settings and reducing interobserver variability that has long challenged endoscopic interpretation.
Currently, adoption is limited by high bias risk, lack of external validation, and interpretability Still, the trajectory of AI suggests a future where these computer technologies will not only support but also elevate human expertise, reshaping the standards of care of diseases managed by advanced endoscopists.
Dr. Singh, Dr. Colletta, and Dr. Marya are based at the Division of Gastroenterology and Hepatology, UMass Chan Medical School, Worcester, Massachusetts. Dr. Marya is a consultant for Boston Scientific, and has no other disclosures. Dr. Singh and Dr. Colletta have no disclosures.
References
1. Navaneethan U, et al. Comparative effectiveness of biliary brush cytology and intraductal biopsy for detection of malignant biliary strictures: a systematic review and meta-analysis. Gastrointest Endosc. 2015 Jan. doi: 10.1016/j.gie.2014.09.017.
2. Stassen PMC, et al. Diagnostic accuracy and interobserver agreement of digital single-operator cholangioscopy for indeterminate biliary strictures. Gastrointest Endosc 2021 Dec. doi: 10.1016/j.gie.2021.06.027.
3. Marya NB, et al. Identification of patients with malignant biliary strictures using a cholangioscopy-based deep learning artificial intelligence (with video). Gastrointest Endosc. 2023 Feb. doi: 10.1016/j.gie.2022.08.021.
4. Marya NB, et al. Multicenter validation of a cholangioscopy artificial intelligence system for the evaluation of biliary tract disease. Endoscopy. 2025 Aug. doi: 10.1055/a-2650-0789.
5. Marya NB, et al. Utilisation of artificial intelligence for the development of an EUS-convolutional neural network model trained to enhance the diagnosis of autoimmune pancreatitis. Gut. 2021 Jul. doi: 10.1136/gutjnl-2020-322821.
AI in General GI and Endoscopy
BY DENNIS L. SHUNG, MD, MHS, PHD
The practice of gastroenterology is changing, but much of it will be rooted in the same – careful, focused attention on endoscopic procedures, and compassionate, attentive care in clinic. Artificial intelligence (AI), like the Industrial Revolution before, is going to transform our practice. This comes with upsides and downsides, and highlights the need for strong leadership from our societies to safeguard the technology for practitioners and patients.
What are the upsides?
AI has the potential to serve as a second set of eyes in detecting colon polyps, increasing the adenoma detection rate (ADR).1 AI can be applied to all areas of the gastrointestinal tract, providing digital biopsies, guiding resection, and ensuring quality, which are all now possible with powerful new endoscopy foundation models, such as GastroNet-5M.2
Additionally. the advent of automating the collection of data into reports may herald the end of our days as data entry clerks. Generative AI also has the potential to give us all the best information at our fingertips, suggesting guideline-based care, providing the most up to date evidence, and guiding the differential diagnosis. The potential for patient-facing AI systems could lead to better health literacy, more meaningful engagement, and improved patient satisfaction.3
What are the downsides?
For endoscopy, AI cannot make up for poor technique to ensure adequate mucosal exposure by the endoscopist, and an increase in AI-supported ADR does not yet convincingly translate into concrete gains in colorectal cancer-related mortality. For the foreseeable future, AI cannot make a connection with the patient in front of us, which is critical in diagnosing and treating patients.
Currently, AI appears to worsen loneliness4, and does not necessarily deepen the bonds or provide the positive touch that can heal, and which for many of us, was the reason we became physicians. Finally, as information proliferates, the information risk to patients and providers is growing – in the future, trusted sources to monitor, curate, and guide AI will be ever more important.
Black Swans
As AI begins to mature, there are risks that lurk beneath the surface. When regulatory bodies begin to look at AI-assisted diagnostics or therapeutics as the new standard of care, reimbursement models may adjust, and providers may be left behind. The rapid proliferation and haphazard adoption of AI could lead to overdependence and deskilling or result in weird and as yet unknown errors that are difficult to troubleshoot.
What is the role of the GI societies?
Specialty societies like AGA are taking leadership roles in determining the bounds of where AIs may tread, not just in providing information to their membership but also in digesting evidence and synthesizing recommendations. Societies must balance the real promise of AI in endoscopy with the practice realities for members, and provide living guidelines that reflect the consensus of members regarding scope of practice with the ability to update as new data become available.5
Societies also have a role as advocates for safety, taking ownership of high-quality content to prevent misinformation. AGA recently announced the development of a chat interface that will be focused on providing its members the highest quality information, and serve as a portal to identify and respond to its members’ information needs. By staying united rather than fragmenting, societies can maintain bounds to protect its members and their patients and advance areas where there is clinical need, together.
Dr. Shung is senior associate consultant, Division of Gastroenterology and Hepatology, and director of clinical generative artificial intelligence and informatics, Department of Medicine, at Mayo Clinic Rochester, Minnesota. He has no disclosures in regard to this article.
References
1. Soleymanjahi S, et al. Artificial Intelligence-Assisted Colonoscopy for Polyp Detection : A Systematic Review and Meta-analysis. Ann Intern Med. 2024 Dec. doi:10.7326/annals-24-00981.
2. Jong MR, et al. GastroNet-5M: A Multicenter Dataset for Developing Foundation Models in Gastrointestinal Endoscopy. Gastroenterology. 2025 Jul. doi: 10.1053/j.gastro.2025.07.030.
3. Soroush A, et al. Generative Artificial Intelligence in Clinical Medicine and Impact on Gastroenterology. Gastroenterology. 2025 Aug. doi: 10.1053/j.gastro.2025.03.038.
4. Mengying Fang C, et al. How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study. arXiv e-prints. 2025 Mar. doi: 10.48550/arXiv.2503.17473.
5. Sultan S, et al. AGA Living Clinical Practice Guideline on Computer-Aided Detection-Assisted Colonoscopy. Gastroenterology. 2025 Apr. doi:10.1053/j.gastro.2025.01.002.