Artificial intelligence in gastrointestinal surgery: A systematic review

Table 1 The 45 studies’ journals included in this systematic review and the number of publications per journal

Journal title	Number of publications
Surgical Endoscopy and Other Interventional Techniques	14
Endoscopy	11
Digestive Endoscopy	7
Annals of Surgery	1
ANZ Journal of Surgery	1
Cirugía y Cirujanos	1
Colorectal Disease	1
Diseases of the Colon & Rectum	1
International Journal of Computer Assisted Radiology and Surgery	1
International Journal of Surgery	1
Journal of Gastrointestinal Surgery	1
Journal of Surgical Research	1
Obesity Surgery	1
Surgery	1
Surgical Laparoscopy Endoscopy & Percutaneous Techniques	1

Table 2 Overview of studies focusing on preoperative planning and risk prediction in gastrointestinal surgery

Ref.	Focus area	AI method/model	Key outcome	Number of data points	Results
Galvis-García et al[13], 2023	Colorectal polyp detection and classification	Deep learning, CNN, CAD (CADe/CADx)	Increased adenoma and polyp detection rates using AI assisted colonoscopy in real time	1038 patients (RCT), 8641 images (CNN study), 466 polyps (CADx), 238 lesions (endocytoscopy)	Sensitivity up to 96.5%, specificity up to 93%, accuracy up to 96.4%, F1 score approximately 94%, NPV up to 99.6%
Zhang et al[14], 2023	Diagnosis of choledocholithiasis in gallstone patients	Machine learning (7 models), AI (ModelArts)	Developed and validated AI model with high diagnostic accuracy for CBD stones prediction	1199 patients (681 with CBD stones)	ModelArts AI: Accuracy 0.97, recall 097, precision 0.971, F1 score 097; machine learning AUCs: 0.77-0.81
Ahmad et al[15], 2022	Detection of subtle and advanced colorectal neoplasia	Deep learning (ResNet 101 CNN)	High sensitivity for detecting flat lesions, sessile serrated lesions, and advanced colorectal polyps	173 polyps, 35114 polyp positive frames, 634988 polyp negative frames across multiple datasets	Per polyp sensitivity: 100%, 98.9%, and 79.5% (in subtle set); F1 score: Up to 87.9%; CNN outperformed expert and trainee endoscopists in detection speed and accuracy
Lei et al[16], 2023	Polyp detection via colon capsule endoscopy	Deep learning (CNN, AiSPEED™)	Feasibility and diagnostic accuracy of AI assisted reading compared to clinician interpretation	Target: 674 patients (597 needed for power; both prospective and retrospective recruitment)	Sensitivity/specificity of AI to be compared to clinician standard; exact results pending study is ongoing
Eckhoff et al[17], 2023	Surgical phase recognition for Ivor-Lewis esophagectomy	CNN + LSTM (TEsoNet), transfer learning	Demonstrated feasibility of knowledge transfer from sleeve gastrectomy to esophagectomy with moderate accuracy	60 sleeve gastrectomy videos, 40 esophagectomy videos (used in combinations across 5 experiments)	Single procedure accuracy: 87.7% (sleeve). Transfer learning: 23.4% overall (4 overlapping phases: 58.6%). Co training max accuracy: 40.8%
Blum et al[18], 2024	Prediction of choledocholithiasis	Logistic regression, RF, XGBoost, KNN; ensemble model	Machine learning models can outperform ASGE guidelines in predicting choledocholithiasis risk using pre MRCP data	222 patients	AUROC: 0.83 (RF), accuracy: 0.81 (ensemble), sensitivity: Up to 0.94, F1 score: Up to 0.82
Axon[19], 2020	Evolution and future of digestive endoscopy	Conceptual AI (future prediction)	Highlights AI’s potential to surpass expert level diagnostic accuracy and support real time treatment decisions		Predicts that AI will revolutionize diagnosis and treatment in endoscopy with real time support tools
Hsu et al[20], 2023	Predicting postoperative GIB after bariatric surgery	RF, XGBoost, NNs	Machine learning models outperform logistic regression in predicting GIB, aiding clinical decision making	159959 patients (632 with GIB)	RF AUROC: 0.764, sensitivity: 75.4%, specificity: 70.0%; XGBoost AUROC: 0.746; NN AUROC: 0.741; logistic regression AUROC: 0.709
Athanasiadis et al[21], 2025	Accuracy of self-assessment in laparoscopic cholecystectomy (CVS quality)		Surgeons frequently overestimate CVS performance; self-assessment alone is insufficient	25 surgeons enrolled, 13 submitted 1 video, 4 submitted 2 videos	No surgeon achieved adequate CVS per expert review; significant discrepancy between self and expert ratings on Strasberg scale
Bisschops et al[22], 2019	Advanced endoscopic imaging for colorectal neoplasia		Guidance on when to use advanced imaging (HD WLE, CE, NBI, etc.) for detection and differentiation of colorectal lesions		AI suggested for future use if validated; various imaging techniques reviewed for ADR, miss rates, lesion detection
Han et al[23], 2025	Intraoperative recognition of PAN during total mesorectal excision	DeepLabv3+ with ResNet50 backbone	AI model (AINS) achieved real time neuro recognition of PAN, aiding nerve preservation during rectal cancer surgery	1780 images (1424 training, 356 validation)	Accuracy: 0.9609, precision: 0.7494, recall: 0.6587, F1 score: 0.7011; AI outperformed surgeons (F1: 0.4568) and operated faster (3 minutes vs 25 minutes)

AI: Artificial intelligence; CNN: Convolutional neural network; CAD: Computer-aided diagnosis; RCT: Randomized controlled trial; NPV: Negative predictive value; CBD: Common bile duct; AUC: Area under the curve; LSTM: Long short-term memory; RF: Random forest; XGBoost: Extreme gradient boosting; KNN: K-nearest neighbors; ASGE: American Society for Gastrointestinal Endoscopy; MRCP: Magnetic resonance cholangiopancreatography; AUROC: Area under the receiver operating characteristic curve; NN: Neural network; GIB: Gastrointestinal bleeding; CVS: Clinical vignette scoring; HD WLE: High-definition white light endoscopy; CE: Chromoendoscopy; NBI: Narrow band imaging; ADR: Adverse drug reaction; PAN: Pelvic autonomic nerves.

Table 3 Summary of studies related to artificial intelligence-based intraoperative guidance in gastrointestinal surgery

Ref.	Surgical context	AI method/model	Guidance function	Data type/source	Real time capability	Performance metrics
Sato et al[24], 2022	Thoracoscopic esophagectomy	DeepLabv3+ (CNN based semantic segmentation)	Recurrent laryngeal nerve identification and navigation	3000 annotated intraoperative images from 20 videos (train/val) + 40 test images from 8 other videos	Yes	Dice: AI 0.58, expert 0.62, general surgeons 0.47
Niikura et al[25], 2022	Upper GI endoscopy	Single shot multibox detector (CNN)	Detection of gastric cancer in endoscopic images	23892 white light images from 500 patients (1:1 matched to expert endoscopists)	No	Per patient: 100%, per image: 99.87%, IOU: 0.842
Yang et al[26], 2022	EUS for subepithelial lesions	ResNet 50 (CNN based deep learning)	Differentiation of gastrointestinal stromal tumors and leiomyomas using EUS	10439 EUS images from 752 patients (multicenter); 132 prospective patients for clinical validation	Yes	AUC: 0.986 (internal), 0.642 (external), accuracy: 96.2%
Schnelldorfer et al[27], 2024	Staging laparoscopy	YOLOv5 (detection), ensemble ResNet18 CNN (classification)	Identification and classification of peritoneal surface metastases	4287 lesions from 132 patients (365 biopsied lesions; 3650 image patches)	No	AUC PR: 0.69, AUROC: 0.78, accuracy: 78%
Guo et al[28], 2021	GI endoscopy (multi lesion)	Deep CNN (ResNet 50 with TTA and transfer learning)	Detecting four lesion categories to support endoscopic diagnosis	327121 WLI images for training; 33959 images in validation from 1734 cases	No (0.05 seconds/image)	Sensitivity: 88.3%, specificity: 90.3%, accuracy: 89.7%
Houwen et al[29], 2022	Colonoscopy		Defines competence standards (SODA) for AI or endoscopists in optical diagnosis	Simulation studies + systematic literature review; ESGE Delphi consensus panel	Yes	Sensitivity ≥ 90%, specificity ≥ 80% (leave in situ); ≥ 80% both (resect and discard)
Tatar and Çubukçu[30], 2024	Colonoscopy	YOLOv8 based CNN (ColoNet)	Identification of neoplastic/premalignant/malignant lesions; biopsy decision support	1760 colonoscopy images (306 patients) for training/validation; 91 external images for real time testing	Yes	mAP50: 0.832, accuracy: 82.4%, sensitivity: 70.7%, specificity: 92.0%

AI: Artificial intelligence; CNN: Convolutional neural network; GI: Gastrointestinal; IOU: Intersection over union; EUS: Endoscopic ultrasonography; AUC: Area under the curve; AUC PR: Area under the precision-recall curve; AUROC: Area under the receiver operating characteristic curve; TTA: Test time augmentation; WLI: White light imaging; SODA: Simple optical diagnosis accuracy; ESGE: European Society of Gastrointestinal Endoscopy; mAP: Mean average precision.

Table 4 Artificial intelligence applications for postoperative monitoring and complication prediction in gastrointestinal care

Ref.	Complication targeted	AI method/model	Timing of prediction	Data source/type	Performance metrics	Clinical utility
van de Sande et al[31], 2022	Need for hospital specific interventions (e.g., reoperation, radiological intervention, IV antibiotics)	Random forest (4 variants tested)	After second postoperative day	EHR data from 3 non-academic hospitals; 18 perioperative variables (e.g., age, BMI, ASA, meds, surgery time)	AUROC: 0.83, sensitivity: 77.9%, specificity: 79.2%, PPV: 61.6%, NPV: 89.3%	Supports early safe discharge and capacity management
Choi et al[32], 2024	Missed small bowel lesions in negative capsule endoscopy	CNN	After initial human reading of SBCE videos	103 negative SBCE videos retrospectively reanalyzed; images from two academic hospitals	CNN detected additional lesions in 61.2% of cases; model had > 96% accuracy in prior study (AUROC = 0.9957)	Reduces diagnostic oversight; changed diagnosis in 10.3%
Blum et al[18], 2024	Choledocholithiasis	Logistic regression, random forest, XGBoost, KNN, ensemble	Before MRCP, using pre intervention data	Retrospective data from 222 patients (clinical, biochemical, imaging variables) from Royal Hobart Hospital	Ensemble & random forest model (accuracy: 0.81, AUROC: 0.83, sensitivity: 0.94, specificity: 0.69, F1: 0.82)	Avoids unnecessary MRCP, triages patients for ERCP
Haak et al[33], 2022	Incomplete response after CRT in rectal cancer	CNNs (EfficientNet B2, Xception, etc.); FFN for clinical data; combined model	After chemoradiation, pre surgery	226 patients; 731 endoscopic images; clinical features from single institute retrospective cohort	EfficientNet B2-AUC: 0.83, accuracy: 0.75, sensitivity: 0.77, specificity: 0.75, PPV: 0.74, NPV: 0.77	Identifies candidates for non-surgical follow up
Noar and Khan[34], 2023	GP	AI derived GMAT threshold using multivariate regression	Pre-treatment (based on GMA from EGG + WLST)	30 patients with GP; GMA via EGG; WLST; gastric emptying tests; symptom scores (GCSI DD, Leeds)	Sensitivity: 96%, specificity: 75%, accuracy: 93%, AUC: 0.95 for GMAT ≥ 0.59	Guides selection for balloon dilation; personalized therapy

AI: Artificial intelligence; IV: Intravenous; EHR: Electronic health record; BMI: Body mass index; ASA: American Society of Anesthesiologists; AUROC: Area under the receiver operating characteristic curve; PPV: Positive predictive value; NPV: Negative predictive value; CNN: Convolutional neural network; SBCE: Small bowel capsule endoscopy; XGBoost: Extreme gradient boosting; KNN: K-nearest neighbors; MRCP: Magnetic resonance cholangiopancreatography; ERCP: Endoscopic retrograde cholangiopancreatography; CRT: Chemoradiotherapy; FFN: Feed forward neural network; AUC: Area under the curve; GP: Gastroparesis; GMAT: Gastric myoelectric activity threshold; EGG: Electrogastrography; WLST: Water loaded stomach test; GCSI DD: Gastroparesis cardinal symptom index-daily diary.

Table 5 Artificial intelligence applications in endoscopic diagnosis and quality control for gastrointestinal cancers

Ref.	Cancer type/lesion	Endoscopic modality	AI method/model	Task/objective	Dataset size/source	Performance metrics	Clinical relevance/impact
Messmann et al[35], 2022	BERN	Upper GI endoscopy	AI assisted deep learning systems	Real time detection and localization of Barrett’s neoplasia	Meta analyzer and real time work (n > 1000 images)	Sensitivity: 83.7%-95.4%, accuracy: 88%-96%	Improves detection of subtle lesions; supports targeted biopsies over Seattle protocol
Choi et al[36], 2022		EGD	CNN (squeeze and excitation network)	Classification of anatomical landmarks and completeness of photo documentation	2599 images from 250 EGD procedures (Korea University Hospital)	Landmark classification: Accuracy 97.58%, sensitivity 97.42%, specificity 99.66%; completeness detection: Accuracy 89.20%, specificity 100%	Enhances quality control in EGD by verifying complete anatomical documentation automatically
Inaba et al[37], 2024		Colonoscopy (preparation phase)	MobileNetV3 based CNN (smartphone app)	AI based stool image classification to assess bowel preparation quality	1689 images from 121 patients; 106 patient prospective validation	Accuracy: 90.2% (grade 1), 65.0% (grade 2), 89.3% (grade 3); BBPS ≥ 6 in 99.0% of app users	Improved bowel prep monitoring; 100% cecal intubation; reduced burden on patients and nurses
Zhang et al[14], 2023	Suspected choledocholithiasis	Not applicable (pre-endoscopy prediction)	ModelArts AI platform (Huawei); 7 machine learning models also tested	Predictive classification of CBD stones before cholecystectomy	1199 patients with symptomatic gallstones; retrospective, single center	ModelArts AI: Accuracy 0.97, recall 097, precision 0.971, F1 score 0.97	May outperform guideline based risk stratification; reduces unnecessary ERCP
Wu et al[38], 2021	EGC	EGD	ENDOANGEL system (CNNs + deep reinforcement learning)	Real time monitoring of blind spots and detection of EGC	1050 patients in multicenter RCT; 196 gastric lesions biopsied	Accuracy: 84.7%, sensitivity: 100%, specificity: 84.3%	Reduced blind spots, improved EGD quality, potential for real time EGC detection in clinical setting
Rondonotti et al[39], 2023	DRSPs ≤ 5 mm	Colonoscopy with blue light imaging	CAD EYE (Fujifilm, Tokyo, Japan), CNN based real time system	Optical diagnosis to support “resect and discard” strategy	596 DRSPs in 389 patients, 4 center prospective study (Italy)	NPV: 91.0%, sensitivity: 88.6%, specificity: 88.1%, accuracy: 88.4%	Meets ASGE PIVI thresholds; may enable safe omission of histology in DRSPs, especially beneficial for nonexperts
Koh et al[40], 2023	Colonic adenomas including SSA	Colonoscopy	GI Genius™ (CADe system, Medtronic, MN, United States)	Real time detection of colonic polyps and ADR improvement	298 colonoscopies; 487 AI “hits”; 250 polyps removed	Post AI ADR: 30.4% vs baseline 243% (P = 0.02); SSA rate: 5.6%	Enhanced ADR even in experienced endoscopists; improved SSA detection; supports AI use in routine colonoscopy
Yuan et al[41], 2022	Gastric lesions (EGC, AGC, SMT, polyp, PU, erosion)	White light endoscopy	YOLO based DCNN model	Multiclass diagnosis of six gastric lesions + lesion free mucosa	31388 images (29809 train/1579 test) from 9443 patients	Overall accuracy: 85.7%; EGC: Sensitivity 59.2%, specificity 99.3%; AGC: Sensitivity 100%, specificity 98.1%	Comparable to senior endoscopists; improved diagnostic accuracy and efficiency; potential for real time support in diverse gastric lesion detection
Munir et al[42], 2024		Not applicable (survey based assessment)	ChatGPT	Evaluation of AI responses to perioperative GI surgery questions	1080 responses assessed by 45 surgeons	Majority graded “fair” or “good” (57.6%); highest “very good/excellent” rate for cholecystectomy (45.3%)	ChatGPT may aid in patient education, but only 20% deemed it accurate; limited utility in reducing message load
Sudarevic et al[43], 2023	Colorectal polyps	Colonoscopy	Poseidon system (EndoMind + waterjet based AI)	AI based in situ measurement of polyp size using waterjet as reference	28 polyps in silicone model + 29 polyps in routine colonoscopies	Median error: Poseidon 7.4% (model), 7.7% (clinical); visual: 25.1%/22.1%; forceps: 20.0%	Significantly improved sizing accuracy; does not require additional tools; useful for clinical polyp surveillance and resection decisions
Tsuboi et al[44], 2020	Small bowel angioectasia	Capsule endoscopy (PillCam SB2/SB3, Medtronic, MN, United States)	CNN (single shot multibox detector)	Automatic detection of angioectasia in CE images	2237 training images, 10488 validation images (488 angioectasia, 10000 normal)	AUC: 0.998; sensitivity: 98.8%, specificity: 98.4%, PPV: 75.4%, NPV: 99.9%	Enables high accuracy detection of angioectasia; may reduce oversight and physician workload during capsule reading
Chang et al[45], 2022		Upper GI endoscopy (EGD)	ResNeSt deep learning model	Evaluate photodocumentation completeness via anatomical classification	15305 training images; 15723 test images from 472 EGD cases	Accuracy: 96.64% (deep learning model), Photodocumentation rate: 78% (esophagus duodenum), 53.8% (pharynx duodenum)	Enables automated auditing of image completeness; higher completeness linked to higher ADR; applicable for routine EGD quality control
Hwang et al[46], 2021	Small bowel hemorrhagic and ulcerative lesions	CE	VGGNet based CNN + Grad CAM	Classification and localization of hemorrhagic vs ulcerative lesions	30224 abnormal + 30224 normal images (train); 5760 images (validation)	Combined model: Accuracy 96.83%, sensitivity 97.61%, specificity 96.04%, AUROC approximately 0.996	Enhanced lesion localization without manual annotation; Grad CAM improves interpretability; supports efficient clinical CE analysis
Jazi et al[47], 2023		Not applicable (survey + clinical scenarios)	ChatGPT 4 (LLM by OpenAI)	Assess alignment of ChatGPT 4 with expert opinions on bariatric surgery suitability and recommendations	10 patient scenarios; 30 international bariatric surgeons	Expert match: 30%; ChatGPT 4 inconsistency: 40%; recommended surgery in 60% vs experts 90%	ChatGPT 4 showed limited alignment and inconsistency; suitable for education, but not yet reliable for clinical decision making
Meinikheim et al[48], 2024	BERN	Upper GI endoscopy (video based)	DeepLabV3+ with ResNet50 backbone (clinical decision support system)	Evaluate add on effect of AI on endoscopist performance in BERN detection	96 videos from 72 patients; 51273 images (train); 22 endoscopists from 12 centers	AI alone: Sensitivity 92.2%, specificity 68.9%, accuracy 81.3%; nonexperts with AI: Sensitivity up from 69.8% to 78.0%, specificity up from 67.3% to 72.7%	AI significantly improved nonexperts’ diagnostic performance and confidence; comparable accuracy to experts; highlights human AI interaction dynamics
Ahmad et al[49], 2021		Colonoscopy		Identify top research priorities for AI implementation in colonoscopy	15 international experts from 9 countries; 3 Delphi rounds	Not performance focused; methodology scores used for consensus	Provides a structured framework to guide future AI implementation research in colonoscopy; emphasizes clinical trial design, data annotation, integration, and regulation
Lazaridis et al[50], 2021		CE		Assess adherence to ESGE guidelines and future perspectives on CE use	217 respondents from 47 countries via ESGE survey	Not model based; survey: 91% performed CE with appropriate indication; 84.1% classified findings as relevant/irrelevant	Highlights variation in guideline adherence; AI identified as top development priority (56.2%); suggests need for standardization and formal CE training
Tian et al[51], 2024		EUS	CNN with attention module	Automatic identification of 14 standard BPS anatomical sites on EUS	6230 training images (1812 patients), internal: 1569 images (47 patients), external: 85322 images (131 patients from 16 centers)	Sensitivity: 89.45%-99.92%, specificity: 93.35%-99.79%, accuracy (internal): 92.1%-100%, kappa: 0.84-0.98	Outperforms beginners, comparable to experts; enables efficient, high quality anatomical identification in EUS; potential for training and standardization
He et al[52], 2020		Upper GI endoscopy (EGD)	CNN models (DenseNet 121, ResNet 50, VGG, etc.)	Automated classification of 11 anatomical sites for quality control and reporting	3704 images from 211 routine EGD cases (Tianjin Medical University Hospital, Tianjin, China)	(DenseNet 121): Accuracy approximately 91.11%, F1 scores up to 94.92% for specific sites	Supports automated quality assurance in EGD via accurate site classification; aids report generation and completeness verification

AI: Artificial intelligence; BERN: Barrett’s esophagus related neoplasia; GI: Gastrointestinal; EGD: Esophagogastroduodenoscopy; CNN: Convolutional neural network; BBPS: Boston bowel preparation scale; CBD: Common bile duct; ERCP: Endoscopic retrograde cholangiopancreatography; EGC: Early gastric cancer; RCT: Randomized controlled trial; DRSPs: Diminutive rectosigmoid polyps; NPV: Negative predictive value; ASGE: American Society for Gastrointestinal Endoscopy; PIVI: Preservation and Incorporation of Valuable endoscopic Innovations; SSA: Sessile serrated adenomas; CADe: Computer-aided detection; ADR: Adverse drug reaction; AGC: Advanced gastric cancer; SMT: Submucosal tumor; PU: Peptic ulcer; DCNN: Deep convolutional neural network; CE: Capsule endoscopy; AUC: Area under the curve; PPV: Positive predictive value; NPV: Negative predictive value; Grad-CAM: Gradient-weighted class activation mapping; AUROC: Area under the receiver operating characteristic curve; LLM: Large language model; ESGE: European Society of Gastrointestinal Endoscopy; EUS: Endoscopic ultrasonography; BPS: Biliopancreatic system.

Table 6 Artificial intelligence based and robotics enhanced surgical education and performance assessment

Ref.	Surgical procedure/task	AI method/model	Assessment modality	Data type/source	Performance metrics	Educational outcome
Garfinkle et al[53], 2022	Gastrointestinal and endoscopic surgery (priority setting)		Survey/Delphi	Survey data from SAGES member surgeons		Identified core needs: Video training, tech adoption
Huo et al[54], 2024	Surgical decision making for GERD	ChatGPT 3.5/4, copilot, Google Bard, Perplexity AI	Prompt based guideline comparison	Standardized clinical vignettes based on SAGES guidelines	Surgeons (accuracy): Bard 6/7 (85.7%), ChatGPT 4 5/7 (71.4%); patients (accuracy): Bard 4/5 (80.0%), ChatGPT 4 3/5 (60.0%); children: Copilot & Bard 3/3 (100.0%)	Revealed inconsistencies in LLM advice; need for medical domain training
Huo et al[55], 2024	Surgical management of GERD	Generic ChatGPT 4 vs customized GPT (GTS)	Prompt based guideline comparison	60 surgeon cases and 40 patient cases based on SAGES and UEG EAES guidelines	GTS (custom GPT): 100% accuracy for both surgeons (60/60) and patients (40/40), generic GPT 4 66.7% (40/60) for surgeons, 47.5% (19/40) for patients	Demonstrated impact of domain customization in LLMs
Nasir et al[56], 2021	Robotic rectal cancer surgery	No AI model used (robotics only)		RCTs, observational studies, registry data	Reduced conversion to open surgery (especially in obese/male patients); improved urogenital function; no difference in long term oncologic outcomes	Reinforced need for structured robotic training (e.g., EARCS)

AI: Artificial intelligence; SAGES: Society of American Gastrointestinal and Endoscopic Surgeons; GERD: Gastroesophageal reflux disease; LLM: Large language model; GTS: Gastroesophageal reflux disease tool for surgery; UEG: United European Gastroenterology; EAES: European Association for Endoscopic Surgery; RCTs: Randomized controlled trials; EARCS: European Academy of Robotic Colorectal Surgery.