Translational artificial intelligence in gastrointestinal and hepatic disorders: Advancing intelligent clinical decision-making for diagnosis, treatment, and prognosis

Abstract

Gastrointestinal and hepatic disorders exhibit significant heterogeneity, characterized by complex and diverse clinical phenotypes. Most lesions present without typical symptoms in their early stages, which poses substantial challenges for early clinical identification and intervention. As an interdisciplinary field at the forefront of technology, artificial intelligence (AI) integrates theoretical innovation, algorithm development, and engineering applications, triggering paradigm shifts within the medical field. Current research trends indicate that AI technology is progressively permeating the entire diagnostic and therapeutic process for gastrointestinal and hepatic disorders, facilitating intelligent transformations in precise lesion detection, optimization of treatment decisions, and prognosis evaluation through the integration of different modal data, construction of intelligent algorithms, and establishment of clinical verification systems. This article systematically reviews the latest advancements in AI technology concerning the diagnosis and treatment of gastrointestinal diseases (such as inflammatory bowel disease and digestive system tumors) and hepatic diseases (including hepato-cirrhosis and liver cancer), emphasizing its application value and transformative potential in critical areas such as imaging omics analysis, endoscopic intelligent identification, and personalized treatment prediction.

Key Words: Gastrointestinal disorders; Hepatic diseases; Artificial intelligence; Diagnosis; Treatment decision; Prognosis

Core Tip: Artificial intelligence (AI) demonstrates transformative potential across gastrointestinal and hepatic disorders. It enhances early detection of subtle lesions (e.g., Barrett's esophagus) by analyzing diverse clinical data, optimizes treatment decisions (e.g., therapy response in liver cancer) via integrated clinical data assessment (including multimodal integration where applicable), and refines prognostic prediction (e.g., recurrence risk in liver cancer). This translational AI enables intelligent clinical decision-making for diagnosis, personalized treatment, and prognosis assessment throughout the patient journey.

INTRODUCTION

Gastrointestinal and hepatic diseases present a considerable challenge to global public health due to their alarmingly high incidence rates, complex clinical manifestations, and heavy economic burden[1]. According to a report by BCC Research, United States, the ed $86.0 billion in 2023 and is projected to surge to $122.5 billion by 2028[2]. Despite this continuous expansion of the global gastrointestinal diagnostics and therapeutics market, the reality is plagued by multiple challenges: (1) Critical diagnostic resource shortages and inequitable distribution: Severe scarcity and uneven allocation of essential resources such as endoscopy lead to diagnostic delays. Moreover, traditional reliance on manual interpretation introduces subjective variability and high miss rates (particularly for early-stage lesions)[3]; (2) Difficulties in personalized treatment: High disease heterogeneity complicates individualized therapeutic decision-making. Risks associated with invasive procedures and steep learning curves limit the widespread adoption of advanced techniques. In addition, inaccurate prediction of treatment response results in ineffective therapies[4]; and (3) Limitations in prognostic models and follow-ups: Existing prognostic models demonstrate insufficient accuracy in predicting recurrence, complications, and survival. Long-term follow-up for high-risk populations suffers from poor adherence and entails substantial costs[5]. The uneven distribution of global healthcare resources further exacerbates these challenges, creating a significant gap in medical treatment, especially in resource-limited settings[6].

In light of these challenges, artificial intelligence (AI) technologies, particularly-machine learning (ML) and deep learning (DL)-are attracting increasing research interest as they offer unprecedented opportunities to revolutionize the approach for addressing core persistent challenges in gastroenterology and hepatology by leveraging their formidable capabilities in data processing, pattern recognition, and prediction. DL models, particularly convolutional neural networks (CNNs), have shown significant potential in medical image analysis, often rivaling or even surpassing the performance of human experts[7-9]. This potential of AI technologies is rapidly translating into diverse practical applications, ranging from enhancing diagnostic accuracy and efficiency (e.g., AI-aided endoscopy, histopathological interpretation, and imaging analysis) and optimizing treatment decision-making (e.g., predicting therapeutic response, assisting surgical planning, and guiding robotic procedures) to enabling prognosis assessment and risk prediction (e.g., developing individualized models for recurrence, complication, and survival prediction). This review aims to systematically delineate the recent advances, proven applications, and future directions of AI in the diagnosis, treatment, and prognosis of gastrointestinal and hepatic diseases.

LITERATURE SEARCHING

Search strategy

We conducted a systematic literature search using the PubMed database for studies published between January 2019 and May 2025. The search strategy employed Boolean operators with the following key terms:

AI-related terms: “Artificial intelligence” OR “AI” OR “deep learning” OR “machine learning” OR “computer aided diagnosis” OR “computational intelligence” OR “convolutional neural network” OR “neural network”.

Disease-related terms: “Barrett’s esophagus” OR “esophageal squamous cell carcinoma” OR “Helicobacter pylori” OR “gastric polyp” OR “gastric cancer” OR “Crohn's disease” OR “small intestinal neuroendocrine neoplasm” OR “ulcerative colitis” OR “colorectal polyp” OR “colorectal cancer” OR “hepato-cirrhosis” OR “hepatic encephalopathy” OR “hepatocellular carcinoma” OR “liver cancer” OR “gastrointestinal disease”.

The final query combined both sets: (AI-related terms) AND (disease-related terms), restricted to titles, abstracts, and keywords of English-language publications.

Inclusion criteria

Studies that met all of the following criteria were included: (1) Population: Focused on AI algorithms for diagnosing, guiding treatment, or predicting outcomes in digestive diseases (esophagus, stomach, intestines, and liver); (2) Study design: Clinical studies (prospective/retrospective cohorts, case-control, and randomized trials) or technical validation studies; (3) Data: Reported explicit performance metrics [e.g., area under the curve (AUC), sensitivity, specificity, and accuracy]; and (4) Publication date: January 2019 or later.

Exclusion criteria

Studies were excluded if they had the following characteristics: (1) Type: Conference abstracts, editorials, case reports, letters, reviews, or non-English publications; (2) Data quality: Lacked key performance metrics or had sample sizes < 50 (except for technical validation studies); (3) Duplication: Were redundant publications (only the most comprehensive dataset was retained); and (4) Relevance: Were purely algorithmic development studies without clinical validation, or animal/cell-based research.

Literature screening and data extraction

Literature screening and data extraction had the following steps: (1) Initial screening: Two investigators (Ren SQ and Chen JM) independently screened titles/abstracts to exclude irrelevant records; (2) Full-text review: Potentially eligible articles underwent full-text assessment against inclusion criteria; and (3) Data extraction: A standardized form was employed to capture: Study design, population characteristics, sample size; AI algorithm type and application (diagnosis/treatment/prognosis); performance metrics (AUC, sensitivity, specificity); and validation method (internal/external). Owing to the widespread practice of incompletely reporting technical details (e.g., specific data augmentation techniques, hyperparameter optimization processes, and detailed validation set splitting strategies) in the original literature - a key challenge to reproducibility in AI research-this review focused on summarizing key parameters wherever feasible (e.g., dataset scale, model architecture, and primary performance metrics). Any discrepancies were resolved through consensus or third-party adjudication.

AI and esophageal diseases

AI and Barrett’s esophagus: Barrett’s esophagus (BE) is the most significant risk factor for the development of esophageal adenocarcinoma (EAC). Consequently, high-definition white-light endoscopy surveillance has become the preferred method for identifying early neoplastic lesions associated with BE[10,11]. Patients with BE require regular endoscopic examinations to detect EAC at a treatable stage[10,11]. However, early Barrett’s neoplasia is frequently missed because lesions are typically flat and exhibit only subtle morphological changes. Furthermore, the low progression rate of BE means most endoscopists encounter these early lesions infrequently in clinical practice. This lack of familiarity with their endoscopic manifestations further increases the risk of missed diagnosis[12,13].

Recent studies have suggested that AI-assisted detection technology can show promising results in response to the aforementioned clinical challenges. De Groof et al[14] employed a prospective approach to conduct a real-time preliminary assessment of a DL-assisted computer-aided detection (CAD) system for identifying BE-associated neoplasia during live endoscopy. This study captured white-light images at 2 cm intervals during endoscopies in 20 patients (10 with non-dysplastic BE and the other 10 with neoplasia) for real-time analysis by using the CAD system. The results demonstrated a system accuracy of 90% (sensitivity 91%, specificity 89%) in a per-level analysis. In other words, the system successfully identified nine out of ten patients with neoplasia with few false positives (only one non-dysplastic patient was misclassified), providing the initial evidence for its high-precision real-time detection capability and supporting its progression to multicenter clinical trials. Similarly, Struyvenberg et al[15] developed and validated a CAD algorithm based on volumetric laser endomicroscopy (VLE) for real-time identification of BE neoplasia. Their prospective multicenter trial enrolled 47 patients and evaluated 318 biopsy-validated VLE targets. The results showed that the algorithm achieved an accuracy of 85% (sensitivity 91%, specificity 82%) on an independent test set, which is statistically superior to the 77% accuracy achieved by ten VLE experts. These studies indicate that AI-assisted technologies can achieve rapid and highly sensitive detection of neoplastic lesions in BE. This holds promise for effectively guiding targeted biopsies, thereby significantly reducing the risk of missed diagnoses associated with traditional random biopsy strategies.

AI and esophageal squamous cell carcinoma: esophageal squamous cell carcinoma (ESCC), the most common type of esophageal cancer, constitutes approximately 80% of global esophageal cancer cases[16]. Its poor prognosis in advanced stages highlights the critical role of early detection and treatment for improving patient outcomes. Although advanced endoscopic techniques such as narrow-band imaging (NBI) and blue light imaging have enhanced detection capabilities, variability in endoscopist experience remains a significant risk factor for missed early lesions. To address this issue, AI has been introduced to augment the diagnostic performance. Waki et al[17] developed a DL system that focused on the screening performance in scenarios with a high missed-diagnosis risk. Trained on 17336 images, this DL system was validated using 100 rigorously simulated videos reflecting rapid non-magnifying endoscopy (50 videos containing ESCC). The DL system achieved a sensitivity of 85.7% (detecting 54 out of 63 Lesions) when the single-frame threshold was ≥ 0.6 and three consecutive positive frames were required for confirmation. However, it had a specificity of only 40%. This difference in specificity primarily stemmed from rapid-pass artifacts specific to fast endoscopic examination, such as those occurring at the esophagogastric junction. Notably, when used as an adjunctive tool, the AI significantly improved the sensitivity of 21 endoscopists from 75.0% to 77.7% (P = 0.00696) while maintaining a high specificity at 91.6%. This result confirms its efficacy in reducing missed diagnoses during screening without increasing the risk of overdiagnosis.

In a methodological contrast, Everson et al[18] aimed for precise stratification of early-stage ESCC. Their clinically interpretable CNN, trained on 67742 magnifying endoscopy with NBI (ME-NBI) images, employed five-fold cross-validation. It predicted the histological status by identifying intrapapillary capillary loop (IPCL) morphology. On an independent test set of 158 images, the model achieved an overall F1 score of 94% (sensitivity 93.7%, specificity 92.4%, AUC 95.8%), which is comparable to the diagnostic performance of expert endoscopists from both Asia and Europe (average F1 scores 96.5%-98%). A key innovation was the integration of class activation map technology, which visualized the decision rationale of the model with the CNN. This integration confirmed the focus of the model on IPCL microvascular structures, which aligned with clinical understanding and enhanced its credibility for clinical application. These differences in performance metrics between the two studies are mainly due to their divergent target scenarios and the use of different validation methodologies. Waki et al[17] prioritized screening sensitivity (85.7%) to address missed diagnoses, utilizing unedited continuous videos simulating real-world screening conditions. However, the lower specificity reported by Waki et al[17] reflects the inherent challenges of this demanding setting. Conversely, the Everson et al’s study, based on static, high-quality ME-NBI images and optimized via cross-validation, aimed to replicate expert-level stratification capabilities for early lesions, achieving strong overall metrics (F1 score 94%, AUC 95.8%)[18]. This complementary nature demonstrates that AI technology can be directionally optimized according to distinct clinical needs-large-scale screening vs precise diagnosis-providing complementary technical support for early ESCC detection. Future prospective clinical trials should aim to definitively establish the real-world clinical utility of these different AI models. The results of key studies evaluating AI applications the diagnosis of esophageal diseases over the past 6 years are summarized in Table 1[19-27].

Table 1 Summary of artificial intelligence in esophageal diseases.

Disease	Application	Ref.	Study design	Region/country	Modality	Test set	AI model	Main findings
BE	Diagnosis	Rosenfeld et al[19]	R	United Kingdom	Questionnaires	1299 patients	ML	ML model with 8 factors (e.g., age) predicts BE (AUC 0.86/0.81), facilitating high-risk screening
	Diagnosis	Abdelrahim et al[20]	P	Europe	WLI	75 patients	CNN	An AI system detected Barrett's neoplasia in real-time endoscopy with 93.8% sensitivity, significantly higher than endoscopists (63.5%)
	Diagnosis	Struyvenberg et al[21]	R	Europe	NBI	157 videos	CNN	Developed a DL-based CAD system for Barrett's neoplasia in NBI videos: 83% accuracy, 85% sensitivity, 83% specificity, processing at 38 fps
	Diagnosis	Hashimoto et al[22]	R	United States	WLI, NBI	1832 images	CNN	AI detects Barrett's early neoplasia at 95.4% accuracy, 96.4% sensitivity via CNN, with real-time lesion localization
Esophageal carcinoma, ESCC	Diagnosis	Tokai et al[23]	R	Japan	WLI, NBI	2042 images	CNN	AI outperformed 13 endoscopists in assessing ESCC invasion depth (accuracy: 80.9%; AUC 0.7873), demonstrating superior diagnostic capability
	Diagnosis	Fukuda et al[24]	R	Japan	NBI, BLI	28333 images	CNN	AI outperformed endoscopists in ESCC detection sensitivity (91% vs 79%) and characterization accuracy (88% vs 75%)
	Diagnosis	Li et al[25]	R	China	WLI, NBI	759 patients	DL	CAD-NBI surpasses CAD-WLI in accuracy/specificity for early ESCC (P < 0.05). Endoscopist combination yields optimal diagnosis (94.9% accuracy, 92.4% sensitivity, 96.7% specificity)
	Diagnosis	Guo et al[26]	R	China	NBI	13144 images	DL	This DL model demonstrates high sensitivity (image 98.04%, video 96.1%) and specificity (image 95.03%, video 99.9%) in real-time diagnosis of esophageal precancerous and early SCC
	Diagnosis	Ohmori et al[27]	R	Japan	WLI, NBI/BLI, ME	21597 images	CNN	AI detected ESCC via non-magnifying endoscopy (NBI/BLI) with 100% sensitivity. With magnification, accuracy reached 83%, comparable to expert endoscopists

R: Retrospective; P: Prospective; ML: Machine learning; CNN: Convolutional neural network; DL: Deep learning; WLI: White light imaging; NBI: Narrow band imaging; BLI: Blue laser imaging; ME: Magnifying endoscopy; AUC: Area under curve; CAD: Computer aided design; ESCC: Esophageal squamous cell carcinoma; SCC: Squamous cell carcinoma; AI: Artificial intelligence; BE: Barrett's esophagus.

AI and gastric diseases

AI and Helicobacter pylori infection: Helicobacter pylori (H. pylori) infection is closely associated with the development of gastric cancer (GC). Therefore, eradication of H. pylori is an effective measure against GC[28]. Although endoscopic findings of H. pylori infection may include features such as enlarged gastric folds and the disappearance of regular collecting venules, the overall accuracy of human visual assessment is limited. Routine diagnosis of GC still relies on biopsy and/or breath tests; however, the latter carries a certain risk of false-negative results[29,30]. To enhance the diagnostic performance of tests for GC, AI-assisted tools are being actively developed. Klein et al[31] proposed a DL-based decision support algorithm combining image processing techniques (such as threshold segmentation and morphological operations) with a CNN (VGG architecture) to detect H. pylori in gastric biopsy tissue sections [Giemsa and hematoxylin & eosin (H&E) stained]. Evaluated on 87 samples validated by PCR/immunohistochemistry, this DL-based algorithm demonstrated 100% sensitivity (significantly higher than the 68.4% sensitivity achieved via microscopic diagnosis) and 66.2% specificity, highlighting its potential as a high-sensitivity screening tool for pathological diagnosis.

Yasuda et al[32] developed an explainable AI system based on a support vector machine (SVM) for diagnosing H. pylori infection using linked color imaging (LCI) endoscopic images. Their method first grouped LCI images based on hue and extracted feature values, and then trained an SVM classifier for diagnosis. The results showed an overall accuracy of 87.6%, which was comparable to that of experienced endoscopists (accuracy rates of 89.5%-90.5%) and that of superior to less-experienced physicians.

AI and gastric polyps: Gastric polyps, defined as protuberant lesions originating from the gastric epithelium or submucosa and projecting into the lumen, exhibit diverse malignant potential. Their presence may indicate an increased risk of intestinal or extra-intestinal malignancies. Adenomatous polyps are considered true neoplastic lesions, with a malignant transformation rate ranging from 6% to 47%[33]. Consequently, the timely diagnosis, biopsy, and resection of gastric polyps are crucial. This facilitates the early elimination of adenomatous lesions, thereby reducing the subsequent risk of carcinogenesis[34]. As most gastric polyps are clinically asymptomatic and often remain undetected prior to routine examinations, it becomes crucial to develop highly effective diagnostic tools. In recent years, AI has provided some novel approaches to enhance the detection and differential diagnosis of gastric polyps. He et al[35] developed a DL-based framework, MultiAttentiveScopeNet, for gastroscopic image analysis. By integrating and fusing features from different layers of ResNet50 after channel normalization, this DL framework significantly improved the analysis of complex gastroscopic images. It achieved a high accuracy of 0.9308 in the differential diagnosis of gastric polyps and protruding lesions, outperforming junior endoscopists and providing a preliminary solution to diagnostic challenges in this field.

Chen et al[36] further improved detection performance with an automatic detection algorithm based on Single Shot MultiBox Detector (SSD), modifying the VGG-16 network architecture. Their algorithm achieved a mean Average Precision of 95.74% in detecting polyps within gastrointestinal endoscopic images. This represents a 12.4% improvement over manual detection and a 5.7% increase over Mask R-CNN while operating 8.41 times faster. These results demonstrate that the AI technology can effectively address the high rates of missed diagnoses and misdiagnoses of polyps in gastrointestinal endoscopic images to a significant extent.

AI and GC: GC, often presenting with insidious and non-specific symptoms, is frequently diagnosed at an advanced stage with a poor prognosis, and has become the fourth leading cause of cancer-related deaths globally[37]. Consequently, early detection and appropriate treatment are crucial for reducing GC-related mortality. With the widespread application of endoscopy in GC diagnosis, there is a growing demand in the medical community for the development of non-invasive early diagnostic methods. However, endoscopic diagnosis of early GC demands substantial expertise. AI holds promise for assisting in achieving more accurate diagnoses and improving image interpretation efficiency[38,39].

Feng et al[40] developed a real-time diagnostic system based on a deep CNN (DCNN) aimed at enhancing endoscopists’ diagnostic accuracy for early GC. They collected test sets comprised images (1289 images) and videos (130 patients) to compare the performance of 12 endoscopists with and without the system’s assistance. The results demonstrated that the DCNN achieved AUCs of 0.917 and 0.930 in the image and video tests, respectively. With AI assistance, endoscopists’ diagnostic accuracy significantly improved (that of novices increased from 81.33% to 95.22%, reaching expert levels), and the diagnosis time decreased. This observation robustly validates the effectiveness of the DCNN system in enhancing GC diagnostic performance.

Beyond diagnosis, AI also exhibits substantial potential in treatment decision-making. You et al[41] applied radiomics analysis of dual-energy computed tomography (DECT) iodine maps to differentiate between GC T1/2 and T3/4a stages, thereby informing neoadjuvant chemotherapy decisions. They data collected from 263 patients across multiple centers, extracted radiomics features from DECT iodine maps, and constructed a nomogram model by integrating ML algorithms (such as logistic regression) with clinical parameters. This model demonstrated excellent performance across the training set, internal validation set, and external test cohort, achieving AUC values of 0.821-0.894. Hence, the ML-based auxiliary method can effectively identify T3/4a stage tumors, thereby enhancing the precision of individualized treatment.

Furthermore, in the field of prognostic assessment, Li et al[42] developed and validated an ML model for predicting lymph node metastasis in GC patients. This study trained the model using retrospective data from 500 patients at a Chinese hospital and performed external validation using data from 824 Asian American patients in the United States National Cancer Institute’s Surveillance, Epidemiology, and End Results (SEER) database. Eight algorithms, including logistic regression and artificial neural networks (ANN), were evaluated. The results demonstrated that the ANN afforded the best performance (accuracy 0.722), with the depth of tumor invasion, size, and Lauren classification identified as key predictors. The researchers also developed a corresponding web application. This accurate and reliable model holds clinical utility for guiding decisions regarding surgical extent (e.g., degree of lymphadenectomy) and personalized postoperative management in GC. These advances demonstrate the guiding role of AI in the diagnostic prediction, treatment guidance, and prognostic assessment of GC. Table 2 summarizes the results of key studies from the past 6 years evaluating AI in diagnosing common gastric diseases[39,43-56].

Table 2 Summary of artificial intelligence in gastric diseases.

Disease	Application	Ref.	Study design	Country/region	Modality	Test set	AI model	Main findings
H. pylori infection	Diagnosis	Martin et al[43]	R	United States	Gastric biopsy	406 patients	CNN	DCNNs accurately recognize gastric pathology damage patterns, especially H. pylori gastritis, and serve as effective screening tools
	Diagnosis	Mohan et al[44]	R	China, Japan	WLI, BLI, LCI	-	CNN	For H. pylori infection diagnosis, CNN achieved 87% accuracy, sensitivity, and specificity, comparable to endoscopists (82.9% accuracy)
	Diagnosis	Nakashima et al[45]	R	Japan	LCI, WLI	515 patients	CNN	Developed LCI/DL-based CAD classifying H. pylori infection into uninfected, active, and post-eradication statuses with 84.2%, 82.5%, 79.2% accuracy. Outperforms WLI and matches expert endoscopists
Gastric polyp	Diagnosis	Yuan et al[46]	R	China	WLI	9443 patients	DCNNs	AI achieved 96.2% accuracy and 88.0% sensitivity for gastric polyps. With AI, junior endoscopists' accuracy significantly improved (96.9%→97.6%), matching seniors
	Diagnosis	Cao et al[47]	R	China	Gastroscopic imaging	2270 images	DL	Improved YOLOv3 with feature fusion boosts small polyp detection in gastroscopic images to 91.6% accuracy, resolving complex background interference
GC	Diagnosis	Horiuchi et al[48]	R	Japan	ME-NBI	2828 images	CNN	CNN system distinguishes EGC from gastritis (sensitivity 95.4%, NPV 91.7%, accuracy 85.3%), aiding clinical diagnosis
	Diagnosis	Li et al[39]	P & R	China	ME-NBI	20341 images	CNN	ME-NBI-based CNN achieves 90.91% accuracy for early GC; 91.18% sensitivity (superior to experts), 90.64% specificity (comparable); overall outperforms non-experts
	Diagnosis	Bu et al[49]	P	China	Liquid biopsy	150 samples	ML	Developing NanoFisher for efficient plasma EV isolation, combining metabolomics and machine learning, achieves 92% accuracy in EGC diagnosis
	Treatment	Wang et al[50]	R	China	CECT	244 patients	ML	CT radiomics distinguishes T2 from T3/T4 GC, guiding neoadjuvant chemotherapy selection
	Treatment	Shang et al[51]	R	China	CECT	311 images	DL	A nomogram built from radiomic and DL features via automated spleen segmentation effectively predicts GC serosal invasion, providing a noninvasive tool for surgical planning
	Treatment	Kang et al[52]	R	South Korea	CT, WLI, biopsy	2927 patients	DL	Developed a Transformer-based multimodal AI system integrating endoscopic images and clinical data. Accurately predicts EGC lymph node metastasis risk (AUC 0.908), guiding treatment decisions
	Treatment	Chen et al[53]	R	China	Laparoscopic surgery video	2460 images	DL	Develop AI models to accurately identify perigastric vessels, enhancing safety and reducing bleeding risks in laparoscopic gastrectomy
	Prognosis	Zhang et al[54]	R	China	CT	669 patients	DCNN	Developed CT-based radiomics nomogram integrating radiomics and clinical factors (e.g., CEA) effectively predicts early recurrence in advanced GC preoperatively (AUC 0.806-0.831)
	Prognosis	Dong et al[55]	R	China, Italy	CECT	730 patients	DL	DLRN accurately predicts GC lymph node metastasis number (C-index 0.797-0.822), outperforming clinical staging and correlating significantly with survival
	Prognosis	Huang et al[56]	R	China	CECT	205 patients	ML	Developed a ML nomogram combining clinical factors (T/N stage) and CT radiomics to predict gastric cancer PNI (validation AUC 0.885), aiding prognosis

R: Retrospective; P: Prospective; ML: Machine learning; LCI: Linked color imaging; CT: Computed tomography; CECT: Contrast-enhanced computed tomography; ME-NBI: Magnifying endoscopy with narrow band imaging; EGC: Early gastric cancer; GC: Gastric cancer; DCNN: Deep convolutional neural networks; H. pylori: Helicobacter pylori; CNN: Convolutional neural network; WLI: White light imaging; BLI: Blue laser imaging; CAD: Computer aided design; AI: Artificial intelligence; EV: Extracellular vesicle; AUC: Area under curve; CEA: Carcinoembryonic antigen; DLRN: Deep learning radiomic nomogram; PNI: Prognostic nutritional index; DL: Deep learning.

AI and intestinal diseases

AI and Crohn's disease: Crohn's disease (CD), a heterogeneous chronic inflammatory bowel disease, presents numerous challenges in clinical management, including intermittent flares, the need for medication adjustments, potential surgical risks, and significant psychological burden[57,58]. As with many chronic diseases, accurately predicting the disease trajectory, treatment response, and prognosis of CD is crucial for individualized precision medicine[59]. To this end, researchers are actively developing AI-assisted tools.

In the field of disease progression prediction, Kim et al[60] integrated clinical variables with transcriptome-wide association study PrediXcan models based on population-specific terminal ileum gene expression data from a Korean cohort. They successfully developed a model predicting the progression of CD in Korean patients to the stenosing (B2) or penetrating (B3) phenotype within 24 months. Analyzing 430 patients, the model achieved AUCs of 0.788 and 0.785 for predicting B2 and B3 progression, respectively, and identified key genes such as CCDC154 and PUS7. This result provides a molecular basis for early intervention in high-risk patients.

Regarding treatment decision optimization, Udristoiu et al[61] focused on the assessment of mucosal healing. They developed a DL algorithm (a CNN-LSTM model) using confocal laser endomicroscopy (CLE) images. Analyzing 6205 CLE images (80% training set, 20% test set), their DL algorithm demonstrated excellent performance on the test set (accuracy 95.3%, sensitivity 92.78%, specificity 94.6%, AUC 0.98), effectively distinguishing inflamed mucosa from the normal one. This result provides an objective, quantitative basis for treatment adjustments, such as medication selection.

For the assessment of disease prognosis, Ferreira et al[62] pioneered the application of a CNN-based DL model for automated detection of intestinal ulcers and erosions in images from PillCam™ Crohn’s Capsule (PCC), a novel panoramic capsule endoscopy system. The model demonstrated high sensitivity (90.0%) and specificity (96.0%), achieving the first automated detection of lesions in PCC images and significantly optimizing the monitoring efficiency and diagnostic performance for CD. This study provides an innovative tool for AI-assisted diagnosis of CD. By enhancing the diagnostic efficacy of the PCC system, this model holds promise for further advancing the precision of mucosal healing assessment and disease staging.

AI and small intestinal neuroendocrine neoplasms: Neuroendocrine neoplasms (NENs) originate from neuroendocrine cells and are commonly found in the gastrointestinal tract and pancreas. Small intestinal NENs (SINENs) are relatively rare[63]. According to the SEER database, the incidence of SINENs has increased by approximately sixfold between 1973 and 2012, with an increase of 1.05 per 100000 in the average incidence rate from 2000 to 2012[64]. Blazevic et al[65] combined two approaches for the diagnosis of SINENs: A systematic visual assessment performed by five clinicians, and the construction of an ML model using 1128 radiomic features extracted from the mesenteric mass and the surrounding tissue. The clinicians’ achieved an AUC of 0.62-0.85 (with a low interobserver agreement), while the best-performing radiomics model, built using features from the surrounding mesentery (SM), achieved a mean AUC of 0.77. This performance was comparable to the assessment by a team of multidisciplinary experts (sensitivity 0.64/specificity 0.68). Hence, it can be suggested that models built using radiomic features from the SM can objectively identify high-risk patients, which provides a basis for prophylactic surgical decision-making. Table 3 summarizes the results of key studies from the past 6 years evaluating the application of AI in diagnosing common diseases of the small intestinal[66-74].

Table 3 Summary of artificial intelligence in intestinal diseases.

Disease	Application	Ref.	Study design	Country/region	Modality	Test set	AI model	Main findings
Crohn's disease	Diagnosis	Li et al[66]	R	China	CTE, histopathology	167 patients	ML	A validated CTE-based radiomics model accurately distinguishes moderate-severe from non-mild fibrosis in Crohn's bowel walls, significantly outperforming radiologists' visual assessments
	Diagnosis	Majtner et al[67]	P	Denmark	pan-CE	7744 images	DL	Auto-detects Crohn's ulcers with 98.4% accuracy. Comparable small/Large bowel accuracy (98.5% vs 98.1%). Distinguishes severity (κ = 0.72)
	Diagnosis	Klang et al[68]	R	Israel	CE	27892 images	CNN	DL model detects Crohn's enteric strictures at 93.5% accuracy (AUC 0.989), precisely distinguishing strictures from ulcers (including severe), enabling automated CE diagnosis
	Treatment	Konikoff et al[69]	R	Israel	APCT	101 patients	ML	Developed an ML model using indicators, such as NLR, to predict Crohn's complication risk in emergency CT (AUC 0.774), enabling risk stratification to reduce unnecessary scans
	Treatment	Con et al[70]	R	Australia	Serum biomarkers	146 patients	DL	RNN with serial biomarkers (AUC 0.754) outperforms logistic regression (AUC 0.659) in predicting biochemical remission (CRP < 5 mg/L) at 12 months post-anti-TNF therapy in Crohn's disease
	Prognosis	Ungaro et al[71]	P	United States, Canada	PEA	265 patients	ML	ML model identified 5 plasma protein markers for penetrating (B3) and 4 for stricturing (B2) complications, outperforming traditional models (B3 AUC 0.79)
	Prognosis	Stidham et al[72]	R	United States	Laboratory examination	2809 patients	ML	Machine learning leveraging routine longitudinal lab data predicts Crohn's surgical risk (AUC 0.78)
SINENs	Diagnosis	Kjellman et al[73]	P	Northern Europe	PET/CT, MRI etc.	278 patients	ML	Multi-plasma protein markers with Random Forest boost SI-NETs diagnosis (Sensitivity 89%, Specificity 91%, AUC 0.99)
	Diagnosis	Clift et al[74]	R	United Kingdom	EHR	382 patients	ML	XGBoost using primary care EHRs effectively identifies undiagnosed high-risk SI-NET patients (AUC 0.869)

R: Retrospective; P: Prospective; CTE: Computed tomography enterography; ML: Machine learning; pan-CE: Pan-enteric capsule endoscopy; APCT: Abdominopelvic computed tomography; RNN: Recurrent neural network; PEA: Proximity extension assay; HER: Electronic health records; DL: Deep learning; AUC: Area under curve; CE: Capsule endoscopy; CT: Computed tomography; CRP: C-reactive protein; TNF: Tumor necrosis factor; PET: Positron emission tomography; MRI: Magnetic resonance imaging; SI-NETs: Small intestine neuroendocrine tumours; SINENs: Small intestinal neuroendocrine neoplasms; NLR: Neutrophil-to-lymphocyte ratio.

AI and colorectal diseases

AI and ulcerative colitis: Ulcerative colitis (UC) is a chronic immune-mediated disease characterized by alternating periods of flares and remission[75]. It currently has no cure, requires lifelong medical therapy[76]. It may even progress to multiple organ failure[77]. Endoscopy plays a central role in UC management, and is crucial for establishing the initial diagnosis, assessing disease extent and activity, and identifying complications, serving as an endpoint in clinical trials for novel therapies[78,79]. However, existing UC endoscopic scoring systems exhibit significant heterogeneity and subjectivity, leading to insufficient interobserver agreement[80]. Recent advances in AI technology have substantially enhanced the objectivity and efficiency of endoscopic assessment in UC.

At the diagnostic level, Wang et al[81] developed a DL-based classification model for colonoscopy images for differentiating CD from UC. Their model achieved an accuracy of 93.35% for UC image classification, outperforming that reported by the best-performing physician (92.39%). This result confirms the value of CNNs in improving the diagnostic efficacy of colonoscopy for UC and offering novel approaches for clinical decision support.

Regarding therapeutic decision support, Higuchi et al[82] utilized a ResNet50 CNN to build an automated system for assessing disease severity via colon capsule endoscopy in patients with UC. This system achieved a training accuracy of 99.2% and a validation accuracy of 98.3%. The spatial distribution maps of colonic disease activity generated by the system not only assist physicians in objectively evaluating the treatment efficacy and formulating individualized treatment strategies but also significantly reduce the physician image review time by approximately 80%. In the domain of prognostic assessment, Bhambhvani and Zamora[83] pioneered the exploration of DL for the automated assessment of Mayo Endoscopic Subscore (MES 1, 2, 3) in patients with UC. Using a ResNeXt-101 CNN to process 777 endoscopic images, their model achieved AUC values of 0.96, 0.86, and 0.89 for MES 3, MES 2, and MES 1, respectively, on the test set, with an overall accuracy of 77.2%. This result indicates the effectiveness of the model in performing individual-level MES classification, improving the quality of endoscopic assessment, and holding promise for optimizing prognostic evaluation in UC patients through precise MES grading.

AI and colorectal polyps: Colorectal cancer (CRC) predominantly originates from colorectal polyps, especially the adenomatous ones. Although initially benign, these polyps may undergo malignant transformation over time if left unmanaged[84]. Colonoscopy is widely used as the core method for screening and preventing polyp progression to cancer[85]. However, this technique is highly operator-dependent. Recent clinical studies have revealed that as high as 22%-28% of polyps are missed during colonoscopy examinations[86]. These missed polyps can cause delayed cancer diagnosis, leading to significantly reduced survival rates (which can be as low as 10%)[87].

AI shows promise in reducing polyp-miss rates and enhancing detection capabilities. Ozawa et al[88] developed a CNN-based AI system for the automated detection and classification of colorectal polyps during colonoscopy. Trained on 16418 polyp images using the SSD architecture, their system achieved a detection sensitivity of 92% and an overall classification accuracy of 83% (achieving 97% adenoma recognition rate under white light imaging), along with demonstrated high-speed processing capability (20 millisecond per frame). These characteristics robustly validate the value of their system in assisting early CRC diagnosis and optimizing treatment decisions. Similarly, Song et al[89] constructed a DL-based CAD system that focused on the histological classification of colorectal polyps to support clinical treatment decisions. This CAD system, trained on a dataset containing 12480 NBI close-focus image patches using ResNet-50 and DenseNet-201 models, achieved test accuracies ranging from 81.3% to 82.4% (kappa values 0.614-0.642). Its performance was significantly superior to that of novice endoscopists (accuracy 63.8%-71.8%) and was on par with expert-level performance (accuracy 82.4%-87.3%). These results demonstrate that AI-driven CAD systems can effectively augment the diagnostic capabilities of endoscopists and serve as powerful tools for real-time optimization of treatment decisions.

AI and CRC: CRC, a clinically common malignancy, is of significant concern because of its increasing incidence and high mortality rates. CRC is identified as the second leading cause of cancer-related deaths, with a five-year survival rate of only 30%-40%, primarily owing to postoperative recurrence and metastasis[37]. In China, CRC ranks third in incidence and fifth in mortality among all malignancies. Although highly curable in its early stages and detectable in asymptomatic individuals through screening methods such as fecal tests, sigmoidoscopy, or colonoscopy[90,91], current screening approaches are often limited by factors like high cost, the need for cumbersome bowel preparation [required for endoscopy and computed tomography (CT) colonography], and requiring patients to handle stool samples themselves (e.g., gFOBT, FIT, and multi-target stool DNA tests)[92,93]. Therefore, the ability to more accurately identify high-risk individuals and tailor treatment plans and prognosis predictions based on their needs is crucial for efficient clinical decision-making, as it can effectively reduce both incidence and mortality related to CRC[94,95].

For early identification, Schneider et al[96] developed an ML algorithm based on routine blood test indicators and demographic data to identify CRC risk in a community. Their algorithm achieved a sensitivity of 35.4% (95%CI: 33.8-36.7) and an AUC of 0.78 (95%CI: 0.77-0.78) for predicting CRC within 6 months. In addition, it identified 3% highest-risk individuals [odds ratio (OR) = 17.7], with a particularly high predictive power for proximal cancer (OR = 34.7) and advanced-stage cancer (OR = 24.8). Their algorithm offers a low-cost, accessible complementary strategy for early CRC screening, which is particularly valuable for high-risk groups. For treatment decision-making, Ito et al[97] utilized DL (CNN) to build an endoscopic diagnostic support system distinguishing cT1b-stage CRC (requiring surgery) from other early-stage lesions (cTis/cT1a). Trained on 190 white-light colonoscopy images (from 41 patients) using AlexNet architecture and three-fold cross-validation, the system successfully diagnosed cT1b with a sensitivity of 67.5%, a specificity of 89.0%, an accuracy of 81.2%, and an AUC of 0.871. This result indicates the potential of their system to effectively assist endoscopists in evaluating tumor invasion depth and optimizing treatment decisions. For prognostic assessment, Jiang et al[98] combined CNN and an ML classifier (Gradient Boosting-Colon) to analyze H&E-stained tissue sections. Their model effectively stratified patients into high- vs low-recurrence risk groups [hazard ratio (HR) = 10.273] and good- vs poor-prognosis groups (HR = 10.687), achieving a prediction accuracy of 75%-77%. Their model could economically and non-invasively assist in formulating personalized chemotherapy regimens (e.g., extending treatment cycles for high-risk patients) by directly extracting prognostic information from routine pathological slides, providing a novel tool for precision medicine. Table 4 summarizes the results of key studies from the past 6 years evaluating AI in common colorectal diseases[89,99-125].

Table 4 Summary of artificial intelligence in colorectal diseases.

Disease	Application	Ref.	Study design	Country/region	Modality	Test set	AI model	Main findings
Ulcerative colitis	Diagnosis	Sutton et al[99]	R	Norway	Endoscopy	8000 images	CNN	AI (especially DenseNet121 model) accurately distinguishes UC from non-UC pathology (AUC 0.999) and grades endoscopic activity (mild/severe, AUC 0.90)
	Diagnosis	Lo et al[100]	R	Denmark	WLI	1484 images	CNN	The CNN model achieved 84% accuracy in distinguishing UC endoscopic severity (Mayo score 0-3), significantly outperforming existing models and standardizing clinical assessment
	Diagnosis	Ruan et al[101]	R	China	Colonoscope	1772 patients	CNN	AI model detects UC/CD at 99.1% accuracy vs physicians' 78%-92.2%, enhancing clinical efficiency
	Diagnosis	Gutierrez Becker et al[102]	R	Europe et al	Endoscopy	1672 videos	CNN	This model directly analyzes raw colonoscopy videos, automatically assessing UC severity (MCES) with high accuracy (AUC 0.84-0.85), reducing manual annotation needs
	Treatment	Bossuyt et al[103]	P	Belgium, Japan	Endoscopy	100 patients	ML	The RD algorithm objectively evaluates UC endoscopic and histologic activity, closely correlated with RHI (r = 0.74), and is sensitive for monitoring therapeutic response
	Treatment	Iacucci et al[104]	P	Europe, North America	HD-WLE, VCE	283 patients	CNN	AI system accurately differentiates UC endoscopic activity/remission (AUC 0.94) and predicts histological remission (83% accuracy), comparable to physicians
	Prognosis	Huang et al[105]	R	China	Colonoscope	856 images	DL, ML	DL/ML-CAD diagnoses mucosal healing (MES 0-1) with 94.5% accuracy and complete healing (MES 0) at 89.0%
	Prognosis	Popa et al[106]	R	Romania	Colonoscope	55 patients	ML	ML models accurately predict endoscopic disease activity one year after anti-TNFα therapy in UC patients (90% accuracy in test set, 100% in validation set)
	Prognosis	Takenaka et al[107]	P	Japan	Endoscopy	2012 patients	DNN	DNN achieves 90.1% accuracy for endoscopic remission and 92.9% for histological remission (UC), reducing biopsy needs
	Prognosis	Maeda et al[108]	P	Japan	Endocytoscope, NBI	145 patients	ML	Real-time AI endoscopy predicts relapse risk in UC remission by analyzing mucosal microvessels (AI-Active 28.4% vs AI-Healing 4.9%, P < 0.001)
Colorectal polyps	Diagnosis	Wang et al[109]	R	China	Colonoscope	1600 patients	CNN	Enhanced GAP model achieves > 98% accuracy (TPR > 96%, TNR > 98%) for colon polyp detection with reduced parameters, enabling lightweight yet accurate diagnosis
	Diagnosis	Song et al[89]	P & R	South Korea	NBI	1169 samples	DL	CAD with NBI predicts polyp histology at 81.3%-82.4% accuracy, outperforming junior physicians (63.8-71.8%) and matching experts (82.4-87.3%), enhancing junior diagnostic performance
	Diagnosis	Jin et al[110]	P & R	South Korea	NBI	2450 images	CNN	AI assistance significantly boosts endoscopists' (especially novices') accuracy for small polyps (< 5 mm) (73.8%→85.6%) and reduces time (3.92→3.37 seconds/polyp)
	Diagnosis	Sakamoto et al[111]	R	Japan	WLI, LCI, BLI	1788 images	DL	CADe sensitive > 94% (WLI/LCI), CADx accuracy > 93% (WLI/BLI), rivals expert endoscopists
	Diagnosis	Zachariah et al[112]	R	USA	WLI, NBI	6223 images	CNN	CNN real-time prediction of colorectal polyp pathology meets PIVI standards: 97% adenoma NPV, > 93% surveillance interval concordance
	Treatment	Wickstrøm et al[113]	R	Europe	Colonoscope	912 images	FCNs	CNNs rely on polyp shape/edges for segmentation; error risk rises significantly in uncertain areas. FCNs combine uncertainty with interpretability visualization, helping doctors pinpoint high-risk regions fast
	Treatment	Su et al[114]	P	China	Colonoscope	659 patients	DCNN	AQCS significantly boosts adenoma detection (28.9% vs 16.5%, P < 0.001), polyp detection, and optimizes withdrawal time and bowel prep during colonoscopy
CRC	Diagnosis	Luo et al[115]	P & R	China	Liquid biopsy	3315 patients	ML	Plasma ctDNA methylation markers (e.g., cg10673833) enable early CRC diagnosis (AUC 0.96) and high-risk group screening (Sensitivity 89.7%)
	Diagnosis	Arabameri et al[116]	R	France, USA, Austria	Fecal microbiota analysis	350 patients	ML	Combining GRNN and DBFS (new feature selection) identified 6 key microbial markers (e.g., Clostridium), enabling high-precision CRC detection (AUC 0.911) but insufficient adenoma sensitivity (AUC 0.724)
	Diagnosis	Zeng et al[117]	P	USA	OCT	26000 images	CNN	PR-OCT system using OCT and RetinaNet distinguishes CRC from normal tissue in real-time with high accuracy (sensitivity 100%, specificity 99.7%, AUC 0.998)
	Treatment	Wang et al[118]	R	China	MRI	240 patients	Faster R-CNN	Faster R-CNN detects positive CRM on rectal cancer pre-op high-resolution MRI with 93.2% accuracy (AUC 0.953); 0.2 seconds/image, highly feasible and efficient
	Treatment	Yang et al[119]	R	China	MRI	89 patients	ML	Pre-treatment ADC radiomics predicts locally advanced rectal cancer resistance to neoadjuvant chemoradiotherapy (AUC 0.83/91.3% accuracy)
	Treatment	Fu et al[120]	R	USA	MRI	43 patients	DL	DL-based radiomics significantly outperform handcrafted features in predicting neoadjuvant chemoradiotherapy response for locally advanced rectal cancer (AUC 0.73 vs 0.64)
	Prognosis	Xu et al[121]	R	China	CT, MRI et al	999 patients	ML	GradientBoosting and LightGBM effectively predict stage IV CRC recurrence risk (AUC up to 0.881); key factors: Chemotherapy, age, LogCEA, CEA, anesthesia duration
	Prognosis	Zhao et al[122]	R	China	Clinical data	7205 patients	ML	ML-based NCDB nomogram predicts metastatic rectal cancer 3-year OS (C-index > 0.77, internal/external validation), outperforming prior models
	Prognosis	Reichling et al[123]	R	France	IHC, WSI	1018 patients	ML	DGMuneS integrating tumor-stroma/CD3+/tumor features better predicts stage III colon cancer recurrence than traditional immune scores (C-index 0.601 vs 0.578)
	Prognosis	Skrede et al[124]	R	Norway, United Kingdom	H&E staining	2467 patients	CNN	Developed a DL-based prognostic biomarker (DoMore v1-CRC) using only routine H&E-stained slides, effectively stratifying Stage II/III CRC risk and outperforming existing markers
	Prognosis	Väyrynen et al[125]	P	USA	H&E staining	1504 samples	ML	ML on H&E slides links dense stromal lymphocytes/eosinophils and their peri-tumoral localization to significantly improved CRC-specific survival

R: Retrospective; P: Prospective; RHI: Robarts Histological Index; RD: Red density; HD-WLE: High definition white light endoscopy; VCE: Virtual chromoendoscopy; TPR: True positive rate; TNR: True Negative Rate; PIVI: Preservation and incorporation of valuable endoscopic innovations; NPV: Negative predictive value; FCNs: Fully convolution networks; AQCS: Automatic quality control system; GRNN: Generalized regression neural network; DBFS: Derivative-based feature selection; OCT: Optical coherence tomography; R-CNN: Faster R-convolutional neural network; CRM: Circumferential resection margin; ADC: Apparent diffusion coefficient; OS: Overall survival; WSI: Whole slide imaging; IHC: Immunohistochemistry; H&E: Hematoxylin & eosin; CRC: Colorectal cancer.

AI and hepatic diseases

AI and hepato-cirrhosis: Hepatic disease, recognized as one of the most life-threatening conditions is a significant global burden, as confirmed by World Health Organization reports[126]. Among them, hepato-cirrhosis-end-stage progressive liver fibrosis-has a high incidence and mortality rates in Asian countries[127]. However, as most patients typically do not exhibit significant clinical symptoms of hepatic diseases before they progress to the decompensated stage, cirrhosis remains often underdiagnosed and is not identified in time[128]. Therefore, it is essential to develop efficient methods for early detection of liver fibrosis. At present, liver biopsy remains the gold standard for diagnosing fibrosis. However, as an invasive procedure, liver biopsy carries risks of complications such as hemorrhage, biliary peritonitis, and pneumothorax[129,130]. Consequently, in clinical practice, experts favor non-invasive and easily accessible ultrasonography over CT or MRI for the initial assessment[131]. However, the diagnostic accuracy of ultrasound is susceptible to the experience and skill level of the operator[132]. AI technology, leveraging its powerful brain-like algorithms, has the potential to enhance the accuracy of cirrhosis diagnosis. Sarvestany et al[133] conducted a retrospective cohort study and developed an integrated ML algorithm for detecting advanced liver fibrosis, including cirrhosis. Six algorithms were trained and validated by utilizing liver biopsy data. Key results showed that the algorithm achieved an AUC of 0.870, outperforming traditional serum biomarkers (such as APRI and FIB-4) and delivering 100% interpretable results, with the performance approaching expert-level proficiency. This advance provides an automated tool for the efficient identification of advanced liver fibrosis, enabling early screening of high-risk populations for cirrhosis and facilitating timely clinical intervention.

AI and hepatic encephalopathy: Hepatic encephalopathy (HE) is defined as a syndrome of the central nervous system dysfunction caused by metabolic disturbances arising from severe liver disease[134,135]. Although minimal HE (MHE), the mildest form of HE, lacks identifiable clinical symptoms, it is characterized by various subtle neurocognitive and psychomotor deficits, such as psychomotor slowing, shortened attention span, executive dysfunction, and memory impairment[134]. Consequently, early diagnosis and effective treatment of MHE are crucial for preventing its progression to overt HE and for improving the quality of life. Zhang et al[136] performed an fMRI dynamic functional connectivity analysis along with a multi-layer modularity algorithm. Their study revealed significant abnormalities in node disjointness within higher-order cognitive networks (e.g., the default mode network and ventral attention network) of patients with MHE[136]. Importantly, these abnormalities are closely correlated with a decline in the attention and visual memory functions of patients. Critically, an SVM model based on this node disjointness metric achieved an accuracy of 88.71% in individually distinguishing patients with MHE from those with cirrhosis without HE. This finding suggests that dynamic brain network node disjointness may serve as an effective biomarker for MHE, providing a novel approach for the early diagnosis of HE.

AI and liver cancer: According to the 2020 global cancer statistics, liver cancer ranks as the sixth most common cancer worldwide and is the third leading cause of cancer-related deaths[36]. Clinically encountered liver cancers are predominantly primary, mainly comprising hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma[137]. The liver is also a frequent site for metastatic tumors[138]. Given the significant differences in treatment strategies for different subtypes of liver tumors[139], multiphasic contrast-enhanced computed tomography has emerged as a critical tool for preoperative diagnosis[140].

However, the differential diagnosis of liver cancer remains challenging, and preoperative misdiagnosis can directly lead to errors in therapeutic decision-making. AI is progressively enhancing various aspects of liver cancer diagnosis and prognosis. In early diagnosis, the HnAIM DL model developed by Cheng et al[141], which processes whole-slide images using an ensemble approach, achieved a high AUC of 0.935 in external validation. Its diagnostic performance even surpassed pathologists' assessments of biopsy samples, significantly improving early HCC detection rates and patient risk stratification capabilities. For prognostic assessment of HCC, Qu et al[142] developed a DL model based on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). This model integrates radiomic features with clinical indicators to accurately identify proliferative HCC subtypes and stratify the risk of early recurrence following curative resection. Their retrospective study analyzed data from 355 patients and employed a hierarchical fusion strategy to build a joint predictive model: (1) Deep feature extraction: It utilizes a hybrid 3D DenseNet-Transformer architecture to automatically learn features reflecting tumor heterogeneity (e.g., vascular invasion patterns) from multi-phase DCE-MRI sequences; (2) Multimodal fusion: It combines the selected imaging features with key clinical indicators [e.g., alpha-fetoprotein (AFP) level and HBV status] via feature-level concatenation. Logistic regression assigns modality-specific weights (e.g., imaging features OR = 22.48 vs AFP OR = 2.22), quantifying their relative contribution to predicting the proliferative phenotype; and (3) Non-linear interaction optimization: The transformer’s multi-head self-attention mechanism dynamically captures complex interactions between imaging features and clinical indicators (e.g., the synergistic effect between elevated AFP levels and delayed tumor enhancement). The model achieved AUCs of 0.98, 0.89, and 0.83 on the training set, internal validation set, and external validation set, respectively. This performance significantly surpassed that of models based solely on clinical data (AUC = 0.77) or imaging data (AUC = 0.55). Apart from enhancing the discriminative power, this integration enabled personalized recurrence risk stratification by generating a composite risk score (e.g., DL prediction probability + AFP threshold). This stratification revealed significantly different 2-year recurrence rates between proliferative (56.0%) and non-proliferative (36.1%) subtypes (P < 0.001), providing a quantitative basis for tailoring follow-up strategies and adjuvant treatment decisions.

AI in precision therapy for liver cancer: AI is revolutionizing precision therapy for liver cancer, and has demonstrated significant potential in predicting treatment efficacy for both targeted therapies and immunotherapies. Recent studies have leveraged multimodal data integration and DL models for accurate treatment response prediction and patient stratification, thereby establishing a new paradigm for individualized therapeutic decision-making.

Radiomics-based predictive models have emerged as a research focus in targeted therapy. Xu et al[143] developed a DL radiomic nomogram (DLRN) that automatically segments CT imaging features to predict responses to FOLFOX-based hepatic arterial infusion chemotherapy in advanced HCC (AUC = 0.988). This model integrates radiomic features from arterial/portal venous phases, DL signatures, and clinical variables (e.g., peritumoral arterial enhancement), significantly outperforming conventional clinical models (P < 0.001). Crucially, the DLRN enables effective survival risk stratification, with a median overall survival (OS) of 26.0 months in predicted responders vs 12.3 months in non-responders (P < 0.001). Its fully automated workflow (processing time < 30 seconds per case) enhances the clinical translation efficiency.

Multimodal AI models have achieved breakthroughs in predicting outcomes of combination immunotherapies. Hua et al[144] constructed a CT radiomics model extracting 1223 quantitative features to predict objective response rates to lenvatinib plus programmed cell death 1 inhibitors with interventional therapy in unresectable HCC (training cohort AUC = 0.900). Notably, 100% of the 14 key predictive features comprised Laplacian of Gaussian-filtered and wavelet-transformed parameters, revealing tumor heterogeneity undetectable by visual assessment. The model stratified patients into high- and low-risk groups with significant differences in progression-free survival (HR = 2.347) and OS (HR = 2.592), providing a quantitative basis for therapeutic selection.

ML-based subclassification systems are advancing precision management. Han et al’s CLAM-C model restratified patients with Barcelona Clinic Liver Cancer stage C HCC using XGBoost, achieving superior AUCs (0.800/0.831/0.715) for 6/12/24-month survival prediction vs those achieved using conventional models (P < 0.001)[145]. Patients were classified into four risk tiers, with median OS ranging from 26.5 months (low-risk) to 3.0 months (very-high-risk). Critically, the model guided treatment optimization: Immune checkpoint inhibitors or transarterial therapies significantly outperformed tyrosine kinase inhibitors (TKIs) in low-risk groups (P < 0.05), while atezolizumab-bevacizumab provided superior survival benefit vs TKIs in high-risk subgroups (median OS: 5.9 months vs 4.2 months; P = 0.004).

The results of key studies evaluating AI applications in common liver diseases over the past 6 years are summarized in Table 5[143,144,146-164].

Table 5 Summary of artificial intelligence in hepatic disease.

Disease	Application	Ref.	Study design	Country/region	Modality	Test set	AI model	Main findings
Hepato-cirrhosis	Diagnosis	Rhyou and Yoo[146]	R	South Korea	US	4950 images	DL	A patch-based DL network for ultrasound cirrhosis diagnosis using synthetic image augmentation, achieving 99.95% accuracy, 100% sensitivity, and 99.9% specificity
	Diagnosis	Luetkens et al[147]	R	Germany	MRI	465 patients	CNN	ResNet50 distinguishes alcoholic from nonalcoholic cirrhosis on MRI: AUC 0.82, 75% accuracy
	Diagnosis	Chang et al[148]	R	United States	Liver biopsy, FibroScan etc.	1370 patients	ML	ML models (especially Random Forest) outperform traditional non-invasive tests (e.g., FibroScan, FIB-4) in identifying significant fibrosis and cirrhosis in NAFLD patients
	Diagnosis	Mazumder et al[149]	R	United States	CT, liver biopsy etc.	351 patients	DL, ML	Combining AI-extracted CT radiomics with routine lab data boosts liver cirrhosis prediction accuracy (AUC 0.84-0.85)
	Diagnosis	Guo et al[150]	P	United Kingdom	NMR spectroscopy	64005 patients	ML	A plasma metabolomics and ML-based nomogram accurately predicts 10-year hepatic cirrhosis complication risk (AUC 0.861), outperforming conventional metrics
Hepatic encephalopathy, HE	Diagnosis	Calvo Córdoba et al[151]	P	Spain	VOG	47 patients	SVM	Automated VOG with SVM detects MHE in 7-10 minutes (93% sensitivity/specificity), outperforming PHES (25-40 minutes)
	Diagnosis	Sparacia et al[152]	R	Italy	MRI	124 patients	ML	MRI radiomics with KNN predicts HE presence (76.5% accuracy); MLP predicts HE severity (≥ stage 2, 94.1%), demonstrating potential for HE diagnosis and staging
	Diagnosis	Chen et al[153]	R	China	MRI	53 patients	SVM	SVM model using gray matter volume discriminates cirrhosis patients with/without MHE at 83.02% accuracy
	Treatment	Liu et al[154]	R	China	TIPS	218 patients	ML	Developed logistic regression model (AUC 0.825) accurately predicts OHE post-TIPS
	Treatment	Zhong et al[155]	R	China	TIPS	207 patients	ANN	ANN model accurately predicts post-TIPS OHE (C-index = 0.863), with 15.9% incidence within 3 months, providing a clinical stratification tool
Liver cancer	Diagnosis	Gao et al[156]	R	China	CECT	723 patients	DL	This model effectively distinguishes malignant liver tumors (HCC, ICC, metastases), achieving 72.6% test-set accuracy and improving physician ICC sensitivity by 26.9%
	Diagnosis	Xu et al[157]	R	China	CT	1049 patients	DL	Swin-Transformer model simplifies LI-RADS classification, effectively distinguishes HCC from non-HCC, and enhances diagnostic performance with clinical data
	Diagnosis	Li et al[158]	R	China	DECT	262 patients	DL	Dual-energy CT deep-learning radiomics nomogram noninvasively predicts HCC MTM subtype, outperforming clinical-radiologic models (AUC 0.87-0.91)
	Diagnosis	Ma et al[159]	R	China	CT, MRI	211 patients	ML	CT/MRI radiomics + clinical features (SVM) achieves highest HCC diagnostic accuracy (82.4%), significantly outperforms single-modality models, and distinguishes HCC vs non-HCC
	Treatment	Hua et al[144]	R	China	CECT	151 patients	ML	Developed and validated a pretreatment CT-based radiomics model predicting response and survival outcomes after triple therapy in unresectable HCC to guide clinical decisions
	Treatment	Xu et al[143]	R	China	CECT	458 patients	DL	The model significantly outperforms single models (externally validated AUC 0.896), effectively distinguishing survival differences (P < 0.001), providing a tool for personalized treatment
	Treatment	An et al[160]	R	China	IATs	2959 patients	ML	Developed MLDSM to risk-stratify unresectable HCC patients undergoing transarterial therapies (alone/combined) for 12-month mortality, guiding clinical treatment decision (e.g. TACE, HAIC)
	Prognosis	Cao et al[161]	R	China	CT, MRI et al	466 patients	DL, ML	Developed pre-/post-op dual-phase DeepSurv model predicting HCC recurrence post-liver transplant; outperformed Milan Criteria (C-index 0.765-0.839), guiding individualized surveillance
	Prognosis	Altaf et al[162]	R	Pakistan	CT, MRI etc.	192 patients	CNN	AI model with tumor size, AFP, and grade predicts post-transplant HCC recurrence risk (validation AUC 0.77)
	Prognosis	Dong et al[163]	R	United States	Clinical data	2038 patients	ML	XGBoost model effectively predicts 1-, 3-, and 5-year survival (AUC > 0.7) in AFP-positive HCC patients, outperforming other algorithms, offering a clinical tool for early intervention
	Prognosis	Yan et al[164]	R	China	MRI	285 patients	CNN	Deep learning-based nomogram integrating imaging, MVI, and tumor number significantly outperforms traditional models in predicting early HCC recurrence (AUC 0.949 vs 0.751)

R: Retrospective; P: Prospective; US: Ultrasound; NMR: Nuclear magnetic resonance; VOG: Video-oculographic; SVM: Support vector machine; PHES: Psychometric hepatic encephalopathy score; MLP: Multilayer perceptron; KNN: K-nearest neighbor; TIPS: Transjugular intrahepatic portosystemic shunt; OHE: Overt hepatic encephalopathy; ANN: Artificial neural network; CT: Computed tomography; CECT: Contrast-enhanced computed tomography; ICC: Intrahepatic cholangiocarcinoma; DECT: Duel-energy computed tomography; IATs: Intra-arterial therapies; TACE: Transarterial chemoembolization; HAIC: Hepatic arterial infusion chemotherapy; MVI: Microvascular invasion; HE: Hepatic encephalopathy; HCC: Hepatocellular carcinoma.

DISCUSSION

Although a large body of research confirms the strong clinical potential of AI in processing clinical images, identifying therapeutic targets, and analyzing large datasets, its real-world implementation faces numerous limitations and challenges. A primary concern is the high dependency of AI models on image quality, with performance requiring significant improvement in complex scenarios (e.g., active hemorrhage) and complex lesion recognition[53]. In specific applications such as intestinal endoscopy, AI systems not only need to incorporate bowel preparation assessment but also urgently require enhanced robustness against common interfering factors like bubbles and residue[67].

More critically, insufficient model generalizability constitutes a core barrier, prominently manifested in three dimensions: The first one is the cross-center/device compatibility issues, as illustrated by Krois et al[165]. A DL model for apical lesion detection trained on German dental panoramic radiographs achieved an F1-score of 54.1% in local testing. However, it sharply dropped to 32.7% when validated at an Indian center. Although cross-center mixed training partially improved the generalizability (raising the Indian F1-score to 46.1%), this reveals significant disparities in equipment parameters, image quality, and disease manifestations across regions. Enhancing generalizability necessitates cross-institutional collaborative training, which inevitably encounters data privacy and standardization challenges[124].

The second dimension comprises limitations in the detection of rare diseases. For example, the rarity of pediatric (particularly congenital) cataracts makes it very difficult to obtain high-quality, large-scale datasets for training and validation. A multicenter randomized clinical trial conducted by Lin et al[166] at Zhongshan Ophthalmic Center, China, revealed a critical performance gap. While the model achieved 98.87% diagnostic accuracy under strictly controlled laboratory conditions, the accuracy significantly declined to 87.4% in real-world multicenter clinical applications. This pronounced “lab-to-clinic translation gap” starkly exposes the fragility of AI model generalizability when confronting data-scarce rare diseases.

Third, the lack of standardized and comprehensive reporting of technical details, such as data preprocessing, hyperparameter settings, and validation strategies, significantly contributes to the poor reproducibility of AI models. This deficiency directly impedes the objective assessment of their true generalization capability.

Together, these three fundamental challenges-poor reporting standards, data/condition dependency, and limited generalizability-constitute major barriers hindering the translation of AI models from research into widespread clinical adoption. Addressing these issues requires establishing standardized reporting guidelines, actively promoting multi-center external validation, and developing more robust algorithms.

To address barriers such as hardware costs and lack of specialized personnel in primary care settings, China has pioneered practical solutions including: Cloud-based AI systems (e.g., Wenzhou’s “Health Cloud Imaging Platform” integrating AI-assisted image analysis, covering > 90% of medical institutions and completing nearly 7 million AI-assisted diagnoses) and tele-diagnosis collaboration models within medical consortiums (e.g., Shanxi Province’s AI diagnostic system deployed across 85 township hospitals). These models aim to overcome technical capacity and resource constraints at the primary care level. However, significant implementation constraints persist, as such solutions heavily rely on high-performance hardware and complex network architectures, which severely limit their widespread adoption and sustainable operation in resource-constrained primary healthcare facilities.

Beyond technical challenges, the potential impact of AI on the clinician-patient relationship remains unclear. Regardless of AI-assisted diagnostic outcomes, the ultimate responsibility lies unequivocally with the clinician. We believe that the development of multimodal AI will extend beyond disease detection and diagnosis to encompass clinical decision support, including pathological diagnosis and metastasis prediction. The ultimate goal of multimodal AI is to create a system capable of comprehensively leveraging large-scale clinical data-such as radiological imaging, endoscopic images, histopathology, and genomic/metagenomic information-to enable precise prediction and assessment of disease risk, treatment response, likelihood of recurrence/metastasis, and both short-term and long-term prognoses.

CONCLUSION

This article reviews the current applications of AI in common gastrointestinal and hepatic diseases. Research on the use of AI in this field is progressively transitioning from an exploratory stage towards building a solid evidence base for clinical practice. Therefore, we believe that AI will become routinely integrated into the clinical management of gastrointestinal and hepatic disorders in the future, particularly in key aspects such as early diagnosis, treatment decision support, and prognostic assessment, where AI will undergo continuous refinement and improvement through application.