This systematic review adheres to the rigorous PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines for comprehensive reporting. Our protocol (CRD42024545296) was duly registered with PROSPERO in May 2024, ensuring transparency and reproducibility in our research methodology. All phases of the review-including title and abstract screening, full-text screening, data extraction, assessment of adherence to reporting guidelines, and evaluation of bias and applicability-were independently performed in duplicate by two reviewers. Any disagreements were resolved through discussion with a third independent reviewer.

Studies were selected based on inclusion and exclusion criteria that focus on the utilization of AI for the diagnosis of mucosal lesions in CD via capsule endoscopy. Inclusion criteria included peer-reviewed research articles and clinical trials that detail the development or application of AI algorithms specific to capsule endoscopy in CD patients. Exclusion criteria eliminated studies not involving AI, those not using capsule endoscopy, and studies focusing on diseases other than CD.

The search will be conducted across several databases including PubMed, Cochrane, EMBASE, and Web of science using keywords such as “artificial intelligence,” “capsule endoscopy,” and “Crohn’s disease” (Supplementary material). Initial screening will involve reviewing titles and abstracts for relevance. Subsequently, full texts of selected articles will be examined to confirm eligibility. This process will be undertaken independently by two reviewers, with discrepancies resolved by consensus or by consulting a third expert reviewer.

Data will be extracted using a standardized data collection form developed to capture specific details relevant to the review objectives. Key data points will include: study author(s), publication year, sample size, AI model description (e.g., type of algorithm, training data size), capsule endoscopy findings, and diagnostic accuracy metrics (specificity, sensitivity). Extraction will be conducted by two independent reviewers to ensure accuracy, with a third reviewer available to resolve any discrepancies. Data will be recorded in a structured digital database, which facilitates data management and subsequent analysis.

Eligible studies for inclusion in the review comprised prospective cohort studies, retrospective analyses evaluating the diagnostic accuracy of AI-assisted capsule endoscopy for detecting CD. To qualify, articles needed to report estimates of overall diagnostic accuracy, sensitivity (%), and specificity (%), accompanied by 95% confidence intervals (CIs). There were no restrictions on the size of the studies.

The Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool was utilized to evaluate the methodological quality of the included articles (Whiting et al., 2011). This tool encompasses four domains: patient selection, index test, reference standard, and flow and timing. Each domain was assessed for high, low, or unclear risk of bias, while the first three domains were also evaluated for high, low, or unclear concerns regarding applicability. The first three areas were also evaluated as applicability issues. Each section is classified as having high, low, or unclear risk of bias. Two investigators independently evaluated the retrieved articles for eligibility, resolving any discrepancies through mutual consensus. Review Manager version 5.3 was employed to create the summary figure for the methodological quality assessment.

The synthesis of data in this review will be tailored to quantitative analyses, given the technical and clinical variability in the studies. Quantitative synthesis, where data permits, will include a meta-analysis to aggregate diagnostic performance metrics such as sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and AUC from the receiver operating characteristic (ROC) analysis.

We used the MIDAS module within STATA (version 18) for statistical analysis. True positives (TP), false negatives (FN), false positives (FP), and true negatives (TN) were constructed to 2×2 contingency tables from artificial intelligence diagnostics of CD patients. The bivariate mixed effects regression model is used for the following indicators: merged sensitivity, specificity, PLR, NLR, DOR, and AUC. Firstly, the heterogeneity of the included studies is evaluated by visually examining the merged comprehensive subject operating characteristic (SROC) curves, and asymmetric shapes indicate significant heterogeneity. Use Spearman correlation analysis to test for heterogeneity caused by threshold effects; Cochran’s Q test and I² value test were used to investigate heterogeneity caused by non-threshold effects. If I² < 50%, it can be considered that there is low heterogeneity between research results. In this case, a fixed effects model was used for merging; If I² ≥ 50%, it can be considered that there is high heterogeneity, and a random effects model is used for merging. Explore the sources of heterogeneity through subgroup analysis and meta regression analysis. Use Deek’s funnel plot to evaluate publication bias, where p < 0.1 indicates asymmetric funnel plot. p ≤ 0.05 is considered statistically significant. Simultaneously using Fagan’s column chart to evaluate the role and clinical value of AI assisted systems in the diagnosis of CD.

A total of 155 articles were identified through a search across four electronic databases. Of these, 36 were found to be duplicates and were removed. During the initial screening, which involved a review of titles and abstracts, 106 articles were excluded. The full texts of the remaining 13 articles were thoroughly reviewed. Out of these, five studies were excluded from the final analysis because they did not align with the focus of this systematic review, which is the role of capsule endoscopy in the assessment of CD. Thus, eight studies were included in the final analysis (Figure 1).

All studies assessed the performance of their AI algorithms using an internal validation test dataset, none reported external validation performance. Among the 8 studies diagnosing CD using capsule endoscopy images, a total of 444 patients (353 with CD and 91controls) were identified from 2020 to 2024. Of these studies, 6 were retrospective and 2 were prospective. Most studies developed AI algorithms using deep learning, while three studies employed convolutional neural network algorithms. Almost exclusively, these studies aimed to evaluate the efficacy of the algorithm in CD patients. Klang et al. (2021) selected capsule endoscopy images from 10 CD patients to train a model for detecting intestinal stenosis lesions, while Marin-Santos et al. (2023) utilized capsule endoscopy images from 38 CD patients to identify various lesions in the intestinal mucosa. An exception to this trend is the study conducted by Brodersen et al. (2024), which notably did not exclude patients with ulcerative colitis and cancer. The number of participants in each study displayed a wide range, varying from a minimum of 10 to a maximum of 133 individuals. Moreover, a limitation across these retrospective studies was the absence of reporting the average age and gender of the participants. Most studies established the AI algorithm based on the deep learning. The included studies could be categorized by analysis based on design, AI algorithm, the number of enrolled patients. These characteristics were evaluated as potential sources of heterogeneity through the subgroup analysis and meta-regression. Detailed characteristics of the studies are presented in Table 1.

Regarding the QUADAS-2 tool, the quality assessment results of included studies are summarized in Table 2. Among the 8 studies included in the final analysis, 2 demonstrated a low risk of bias, 6 study at unclear risk. Bias in the included studies primarily stems from the domains of patient selection, index test, flow and timing.

Among the 8 studies of patient-based analysis, the sensitivity (Figure 2a), specificity (Figure 2b), PLR (Figure 2c), NLR (Figure 2d), diagnostic score (Figure 2e), and DOR (Figure 2f) of AI for CD were 0.94 (95% CI, 0.93-0.96), 0.97 (95% CI, 0.95-0.98), 32.7 (95% CI, 19.9-53.6), 0.06 (95% CI, 0.04-0.07), 6.36 (95% CI, 5.69-7.03), and 576 (95% CI, 295-1127), respectively, in Figure 2. The summary receiver operating characteristic (SROC) curve for the included studies (n = 8) was presented in Figure 3, where the calculated area under the curve (AUC) was 0.99 (95% CI, 0.97-0.99). To investigate the clinical utility of AI, a Fagan nomogram was generated. As shown in Figure 4, with a pretest probability of 20%, the post-test probability for a positive result was 89%. Given a negative test result, the negative likelihood ratio of 0.06 reduced the posterior probability to 1%.

The meta-analysis of AI-assisted capsule endoscopy for CD lesion detection revealed significant heterogeneity. To explore the sources of this heterogeneity, we performed meta-regression and subgroup analyses. Results demonstrated that heterogeneity in sensitivity was significantly influenced by design type, sample size, and algorithm type; whereas heterogeneity in specificity was primarily attributed to sample size and algorithm type (Table 3 and Figure 5).

We performed a publication bias analysis for the included studies. No significant bias was observed in diagnostic Crohn’s disease efficacy of AI via capsule endoscopic images. Figure 6 shows the Deek’ funnel plot indicated no evidence of publication bias (p = 0.56).

CD is an autoimmune disorder that affects the entire digestive tract and can progress rapidly (Chapman and Satsangi, 2024). Capsule endoscopy offers significant advantages for diagnosing CD, as it minimizes the need for multiple invasive procedures and reduces patient discomfort (Marquès Camí et al., 2020). However, capsule endoscopy produces a large volume of images, some of which may be blurry or irrelevant. Moreover, gastrointestinal mucosal lesions of CD include various manifestations such as edema, ulcers, fistulas, stenosis, etc., all of which contribute to the complexity and challenge of making an accurate diagnosis (Deng et al., 2024). Especially for inexperienced clinicians in primary hospitals, differentiating CD, UC and non-IBD colitis has remained a dilemma.

Traditionally, endoscopy diagnosis is operator-dependent and subjective, recent studies have demonstrated the substantial potential of AI in medical diagnostics (Khan et al., 2020; Mitsala et al., 2021; Huang et al., 2022). AI-assisted endoscopy can provide a valuable second opinion, potentially reducing operator dependency (Parkash et al., 2022). This approach, characterized as computer-aided diagnosis, significantly improves the diagnostic accuracy of endoscopy. Endoscopist combined with AI analysis will enhance the probability of uncovering crucial gastrointestinal mucosal lesions under endoscopy, thereby advancing the accuracy and effectiveness of diagnostic assessments. Recently, deep learning, as the subset of AI, grown dramatically with a high overall coverage in medical research. Based on the detailed description of Table 1, it is evident that the deep learning algorithm undergoes constantly updates. Thousands of computational iterations, these algorithms can “learn” the unique features of images or data with a given classification, then emulate human problem-solving and identify specific capsule endoscopy images of CD. Klang et al. (2020) presented the results of 5 experiments performed on randomly split images (80% training, 20% testing), ultimately enhancing classification accuracy through the integration of these data. Throughout the validation phase, they demonstrated an enhanced performance, achieving accuracies ranging from 95.4 to 96.7%. Vallée et al. (2020) and De Maissin et al. (2021) employed a multifaceted strategy incorporating a diverse array of deep learning models to secure a thorough assessment of the characteristics of histological images and endoscopic data, yielding an overall F1 score of 91.46% with the VGG16 system. Marin-Santos et al. (2023) demonstrated high recognition accuracy utilizing the convolutional neural network system, exhibiting substantial precision in test phase and achieving a 95–99% sensitivity range.

This article provides a comprehensive review of the diverse applications of AI in diagnosing CD patients through the analysis of capsule endoscopy images. Deep learning exhibits high diagnostic accuracy for CD patients undergoing capsule endoscopy, particularly when utilizing convolutional neural networks. The applications of AI are not only in diagnose CD, but also in prognosis and predicting treatment response. It is essential for managing diseases, mitigating the risk of prolonged complications, and enhancing the quality of life for CD patients. A significant finding of this study is the robustness of the AI algorithm’s diagnostic performance, while studies involving a large patient population and high methodological quality demonstrated higher diagnostic performance, the difference was not substantial.

Despite employing a comprehensive and thorough search strategy, several limitations must be acknowledged regarding both the evidence and the review itself. We were able to identify only 8 eligible studies. Any of these studies presented an unclear risk of bias due to their design. But most studies either had small sample sizes or included only patients with CD, which may have compromised the precision of the effect estimates. The studies included in the meta-analysis lack external data validation, which compromises the reliability of their findings. Most studies utilize deep learning as the AI algorithm, with only 3 studies employing convolutional neural network.

Compared to previous AI-assisted studies in endoscopy, those applications are predominantly utilized to differentiate polyps, tumors, peptic ulcers, and IBD. This review has demonstrated that the majority of the included studies are capable of detecting multiple abnormalities in patients with CD. However, differentiating between IBD subtypes remains a notable gap, particularly due to their complexity and the significant variability observed among individual patients. AI-assisted colonoscopy applications specifically designed for UC patients have been introduced exclusively in Japan. Due to the requirement of specialized endoscopes, this software has not been widely adopted in clinical practice (Ogata et al., 2025). Additionally, it suffers from multiple biases, particularly the unequal distribution of CD patients among the participants, the absence of randomization in selecting subjects, the utilization of datasets from different resources, and the use of more uneven image resolution, all of which restrict the generalizability of the findings. Although these findings employing a multimodal imaging technique which facilitated a more thorough evaluation of histological characteristics, AI applications for assessing architectural modifications and inflammatory infiltrates were less impressive, further advancements are necessary to enhance the accuracy and dependability in identification. This meta-analysis also presents the results of an advanced AI-assisted diagnostic system, but the studies included were unable to detect and analyze CD activity, and estimate the corresponding endoscopic activity. To enhance diagnostic capabilities, AI must be trained and rigorously validated using vast amounts of data, allowing it to learn the intricacies of these diverse manifestations and ultimately improve its diagnostic accuracy. Therefore, more prospective studies focusing on the application of AI in clinical practice, particularly in diagnostic equipment, is essential.