Abstract
Objective
Pan-intestinal capsule endoscopy (PCE) offers a safer, more effective alternative to colonoscopy for detecting potentially haemorrhagic lesions (PHL) in suspected mid-lower gastrointestinal bleeding (MLGIB), though it is limited by time-consuming review and missed lesions. We compared the diagnostic performance of artificial intelligence-assisted PCE (AI-PCE) versus conventional reading PCE (CR-PCE) and colonoscopy.
Methods
We retrospectively analysed 100 prospectively enrolled patients undergoing PCE for suspected MLGIB using an externally validated convolutional neural network. Diagnostic performance of AI-PCE, CR-PCE and colonoscopy was evaluated against a consensus reference standard. Accuracy metrics (sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV)) were assessed overall and by lesion type and intestinal segment.
Results
AI-PCE detected PHL in 60% of patients versus 42% with CR-PCE (p<0.01). Lesions included vascular (51% vs 33%, p<0.01), ulcers/erosions (16% vs 7%, p=0.012), protuberant (5% vs 4%, p=1.0) and active bleeding (7% vs 7%, p=1.0). AI-PCE achieved higher sensitivity than CR-PCE (95% vs 67%, p<0.0001) with comparable specificity (97% vs 97%), PPV (98% vs 98%) and superior NPV (92% vs 63%, p=0.0015). For the small bowel, AI-PCE outperformed CR-PCE in sensitivity (96% vs 59%, p<0.0001) and NPV (97% vs 76%, p=0.0010). In colon, AI-PCE also showed greater sensitivity (90% vs 68%, p=0.027) and NPV (94% vs 86%, p = 0.066). Compared with colonoscopy, AI-PCE was markedly more sensitive (90% vs 32%, p<0.0001) with higher PPV (100% vs 65%, p<0.001) and NPV (94% vs 65%, p<0.0001).
Conclusion
AI-PCE significantly improves diagnostic accuracy over conventional reading and colonoscopy, offering superior sensitivity without compromising specificity, and may establish a new standard for PCE in MLGIB.
Keywords: GASTROINTESTINAL BLEEDING, SMALL BOWEL ENTEROSCOPY, SMALL BOWEL DISEASE, COLONIC DISEASES
WHAT IS ALREADY KNOWN ON THIS TOPIC
Pan-intestinal capsule endoscopy (PCE) is a safe and effective, minimally invasive diagnostic tool for evaluating suspected mid-lower gastrointestinal bleeding (MLGIB). However, conventional reading PCE (CR-PCE) is time-consuming and reader-dependent, with risk of missed lesions. Artificial intelligence (AI) has shown promise in improving lesion detection in small-bowel capsule endoscopy, but evidence in PCE is limited.
WHAT THIS STUDY ADDS
For the first time in full videos, AI-assisted PCE (AI-PCE) was demonstrated to have a significantly higher sensitivity and negative predictive value than CR-PCE for potentially haemorrhagic lesions detection, in the small bowel and colon.
Compared with conventional colonoscopy, AI-PCE achieved markedly greater sensitivity and predictive values, while maintaining high specificity.
The improvements were consistent across key lesion types, including vascular lesions and ulcers/erosions.
HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
AI-PCE offers a more accurate and reliable diagnostic option than CR-PCE or colonoscopy for suspected mid-lower gastrointestinal bleeding. This could redefine the diagnostic approach to MLGIB, positioning capsule endoscopy as a minimally invasive first-line tool, reducing reliance on invasive procedures like colonoscopy.
By reducing false negatives without increasing false positives, AI-PCE has the potential to decrease the need for repeated procedures and improve patient outcomes.
These findings support the integration of AI into capsule endoscopy workflows, although further validation and cost-effectiveness analyses are needed.
Introduction
Capsule endoscopy (CE) was initially developed to provide a minimally invasive, radiation-free and patient-friendly procedure to evaluate the small bowel. Over time, subsequent technological advances introduced colon CE to offer a more detailed evaluation of the colon and, more recently, pan-intestinal CE (PCE), capable of evaluating both the small bowel and colon.1 2
In clinical practice, gastrointestinal bleeding, whether overt (presenting as melena and/or haematochezia) or occult (chronic iron deficiency anaemia), is an indication for the use of CE after a normal upper gastrointestinal endoscopy and conventional colonoscopy.3 The potential role of PCE versus the current standard of care, colonoscopy, as a first-line test to investigate patients with suspected mid-lower gastrointestinal bleeding (MLGIB) is gaining importance with much research in this field.2 This occurs due to the low diagnostic yield of colonoscopy in these cases, suggesting that this research will potentially, in the future, influence guidelines and eventually reduce the need for invasive procedures such as colonoscopy and/or device-assisted enteroscopy, allowing a more patient-friendly and green practice by reserving these examinations for patients that need additional procedures. Although PCE shows promising results in detecting potential bleeding sources, it remains unclear whether artificial intelligence (AI)-assisted analysis substantially improves diagnostic accuracy.
Nevertheless, CE, particularly PCE, is accompanied by some notable limitations. The dual-camera setup produces a large number of frames per examination, substantially increasing the complexity and duration of video analysis, with reported reading times up to 120 min.1 4 Moreover, this process is also susceptible to errors, as the identification of any clinically relevant lesion (ie, diagnostic accuracy) is heavily reliant on the examiner’s clinical expertise,5 and influenced by common drawbacks such as suboptimal bowel preparation, incomplete examinations, as well as the rapid transit or the random dislocation of the capsule, particularly in the colon. These challenges underscore the potential value of integrating AI tools for video-reading assistance, notably convolutional neural networks (CNNs), specifically tailored towards image and pattern recognition.6,8 These AI models have already been integrated into clinical gastroenterology practice, most notably in detecting polyps during colonoscopy.9 10 In addition, there is research demonstrating the role of AI in CE image analysis for a variety of gastrointestinal lesions, including haemorrhagic lesions.11 Therefore, by potentially increasing efficiency and reducing the time required for video analysis while upholding a high level of accuracy (reducing the false negative rate),12 AI-powered PCE can broaden its clinical applications and make this diagnostic modality an earlier option in clinical algorithms.
In addition, PCE poses greater interpretive demands compared with conventional small bowel CE. It uses a dual-camera system that simultaneously captures forward and backward views at higher frame rates and longer video duration, generating a substantially larger number of frames per examination and increased data complexity. Furthermore, the capsule must traverse two anatomically and visually distinct environments - the small bowel and the colon, each with different mucosal patterns, vascular architectures and lesion morphologies. Random capsule movements, incomplete distension and variable bowel preparation quality further complicate interpretation. These factors significantly increase reading time and the likelihood of oversight when relying solely on human readers, making AI integration particularly well suited for PCE.
Given these considerations, our study aimed to determine whether AI-assisted PCE (AI-PCE) improves the detection rate of potentially haemorrhagic lesions (PHL) in the small bowel and colon, compared with the conventional human reading PCE (CR-PCE). As a secondary objective, we evaluated the accuracy of AI-PCE versus conventional colonoscopy, the current standard first-line procedure for MLGIB in clinical practice.
Methods
Study design and patients
This study retrospectively uses data originally collected from a published prospective unicentric study using 100 consecutive patients with suspected MLGIB.2 MLGIB was defined as overt gastrointestinal bleeding (melena and/or haematochezia) and/or IDA (<12.0 g/dL in females and <13.0 g/dL in males) following a non-diagnostic upper endoscopy.
These patients underwent PCE, using double-headed capsules (PillCam Crohn’s, Medtronic) and same day conventional colonoscopy. After non-diagnostic upper endoscopy, bowel preparation was initiated with a split-dose regimen of polyethylene glycol solution (2L+1L, separated by 6–8 hours), along with two sodium phosphate boosters after capsule entered the duodenum and again 3 hours later. Colonoscopies were performed later the same day by independent blinded physicians, using standard techniques, under anaesthesiologist-administered deep sedation and without requiring additional bowel preparation.
The flowchart of study methodology is summarised in figure 1. Exclusion criteria included contraindications for CE (eg, high risk of obstruction), neuropsychiatric or medical comorbidities impeding the protocol compliance, identified cause for bleeding on upper endoscopy, presumed non-gastrointestinal causes for IDA (eg, vaginal bleeding) and pregnancy/breastfeeding. None of the included patients had a history of gastrointestinal surgery resulting in excluded intestinal segments.
Figure 1. Flowchart of study design. CE, capsule endoscopy; MLGIB, mid-lower gastrointestinal bleeding.
Conventional and AI-assisted PCE reading
PCE and colonoscopy findings were reviewed to document the presence or absence of PHL in the small bowel or colon, defined using standard terminology as proposed by the International Capsule Endoscopy Group.13 14 These included vascular lesions, ulcers/erosions, polyps/tumours and active bleeding. All findings were confirmed by experts in CE.
PCE videos were analysed using the latest version of published CNN developed for automated panenteric detection of PHL in the small bowel and colon.15,17 A dual deep learning architecture was implemented, deploying two independent CNNs to account for the distinct anatomical and visual characteristics of different gastrointestinal segments. Specifically, one CNN was exclusively trained and optimised for the analysis of small bowel images, while the second was dedicated to the evaluation of colon segments. Gastroenterologists with CE experience, based at a separate centre and blinded to both the conventional PCE and colonoscopy reports, conducted AI-assisted PCE reading. The process involved uploading the full video into the CNN software, which identified the frames most likely to contain PHL, streamlining the analysis by focusing on relevant frames, instead of the entire video. The CNN-selected frames were then carefully reviewed by a panel of expert gastroenterologists to determine the presence or absence of PHL.
The quality of bowel preparation in the colon was considered adequate in PCE if CC-CLEAR (Colon Capsule CLEansing Assessment and Report) score≥6 and no segmental scores<2,18 and Boston Bowel Preparation Scale score≥6, with no segmental score<2 in colonoscopies.19 The small bowel was considered adequately clean if SB-CLEAR (Small Bowel CLEansing Assessment and Report) score≥6, with no segmental score<2.20
Study endpoint and outcome analysis
The aim of this study was to compare the diagnostic accuracy of CR-PCE with AI-PCE. The gold standard was determined by combining information from CR-PCE and AI-PCE findings.
In concordant cases, where both methods identified the same lesion categories, the findings were directly included in the gold standard. For discordant cases, where lesion categories differed between methods, an independent gastroenterologist reviewed the findings to resolve discrepancies. This element was an expert with more than 10 years’ experience in CE and a record of at least 1000 PCE or small bowel CE readings, not involved in the initial CR-PCE readings or the AI-PCE readings, ensuring independence from both assessment processes. The final gold standard included all lesions confirmed through this process.
Additionally, as a secondary objective, the diagnostic performance of AI-PCE in detecting colonic PHL was compared with conventional colonoscopy using the same metrics.
Diagnostic performance was evaluated on a per-patient basis, with detection defined as identification of at least one PHL of a given category in that patient, by calculating sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) for AI-PCE, CR-PCE and colonoscopy. These measures were calculated both for the entire gastrointestinal tract and separately for the small bowel and colon. In addition, analyses were stratified by lesion category—vascular lesions, ulcers/erosions, haemorrhagic protruding lesions and active bleeding, for both overall and segment-specific datasets. Continuous variables are expressed as mean±SD (minimum - maximum). Categorical variables are presented as counts and percentages. Sensitivity, specificity, PPV and NPV were expressed with absolute counts and proportions. Comparisons between diagnostic methods were performed using the Fisher’s exact test, χ2 test or the McNemar test as appropriate, with a significance defined as p<0.05. The area under the receiver operating characteristic curve was calculated for each diagnostic method, both overall and by lesion category.
Reporting
This study is reported in accordance with the Standards for Reporting Diagnostic Accuracy Studies (STARD) guidelines (see online supplemental file 3 for STARD checklist and online supplemental figure S1) for STARD flow diagram).
Patient and public involvement
None.
Results
Characterisation of study population
100 consecutive patients with suspected MLGIB were included, of which 65% females, with a mean age of 66.5 years (median age 70 (range 18–92) years), presenting with IDA (81%) or overt bleeding (19%). All patients had an upper gastrointestinal endoscopy revealing no PHL. Approximately one-third (34%) of patients were receiving antiplatelet medication, and one-fourth (26%) were receiving anticoagulant drugs. Non-steroidal anti-inflammatory drugs, including aspirin, were used by 37% of patients. Approximately three-quarters (74%) of the patients presenting with overt gastrointestinal bleeding were submitted to PCE and colonoscopy in less than 2 weeks after the bleeding episode. Patient demographics and baseline characteristics are summarised in online supplemental Table S1.
Procedural details
PCE was complete in 76% of cases, while completion rate for colonoscopy was 95% (p<0.001). The quality of bowel preparation in the colon was considered adequate in 72% of PCE examinations, versus 85% in colonoscopies, p=0.01. The small bowel was adequately clean in 96% of patients, with a completion rate of 98% at PCE.
Detection rates
Overall, PCE detected PHL in 63% of patients, including small bowel and/or colon vascular lesions in 54%, ulcers/erosions in 17%, haemorrhagic protruding lesions in 5% and active bleeding in 7% of cases. The distribution of PHL as the gold standard diagnosis per segment is summarised in online supplemental Table S2.
PHL were detected by AI-PCE versus CR-PCE in 60% versus 42% (p<0.01) of patients, including small bowel and colon vascular lesions in 51% versus 33% (p<0.01), ulcers/erosions 16% versus 7% (p=0.012), haemorrhagic protruding lesions 5% versus 4% (p=1.0) and active bleeding 7% versus 7% (p=1.0). Figure 2 illustrates some examples of generated heatmaps of different types of lesions in PCE.
Figure 2. Examples of generated heatmaps of different types of lesions in pan-intestinal capsule endoscopy: (a) small bowel ulcer; (b) small bowel angioectasia; (c) small bowel haemorrhagic protuberant lesion; (d) colon polyp; (e) colon angioectasias with active bleeding; (f) haemorrhoids.
The diagnostic yield of AI-PCE, CR-PCE and colonoscopy is summarised in online supplemental Table S3 and figure 3.
Figure 3. Diagnostic yield of AI-PCE, CR-PCE and colonoscopy per type of lesion. AI-PCE, artificial intelligence-assisted pan-intestinal capsule endoscopy; CR-PCE, conventional reading pan-intestinal capsule endoscopy; PHL, potentially haemorrhagic lesions; PR, haemorrhagic protruding lesions; PUE, ulcers or erosions; PV, vascular lesions; P3, active bleeding.
Considering the detection of PHL in the entire enterocolonic tract, AI-PCE showed a sensitivity of 95% compared with 67% for CR-PCE (p<0.001), with both methods having the same specificity (97%) and PPV (98%), while NPV was 92% for AI-PCE versus 63% for CR-PCE (p=0.0015). Compared with colonoscopy, AI-PCE demonstrated markedly higher sensitivity (95% vs 32%, p<0.0001) and NPV (92% vs 45%, p<0.0001), with comparable specificity. Table 1 summarises the diagnostic accuracy of AI-PCE, CR-PCE and colonoscopy for all enterocolonic PHL, and per lesion type. When analysed by intestinal segment (table 2), AI-PCE showed superior performance in both the small bowel and the colon. In the small bowel, sensitivity was significantly higher with AI-PCE than with CR-PCE (96% vs 59%, p<0.0001) with a corresponding improvement in NPV (97% vs 76%, p=0.0010), while specificity and PPV remained uniformly high in both methods. When stratified by lesion category, AI-PCE outperformed CR-PCE for vascular lesions and ulcers/erosions, with similar performance for haemorrhagic protruding lesions and active bleeding.
Table 1. Diagnostic accuracy of AI-PCE versus CR-PCE versus colonoscopy (all enterocolonic PHL).
| Sensitivity (%) (95% CI) |
Specificity (%) (95% CI) |
PPV (%) (95% CI) |
NPV (%) (95% CI) |
AUROC (%) (95% CI) |
|
|---|---|---|---|---|---|
| PHL (n=63) | |||||
| AI-PCE | 95.2 (86.9 to 98.4) | 97.3 (86.2 to 99.5) | 98.4 (91.3 to 99.7) | 92.3 (79.7 to 97.3) | 0.963 (0.923 to 0.993) |
| CR-PCE | 66.7 (54.4 to 77.1) | 97.3 (86.2 to 99.5) | 97.7 (87.9 to 99.6) | 63.2 (50.2 to 74.5) | 0.820 (0.750 to 0.881) |
| Colonoscopy | 31.7 (21.6 to 44) | 94.6 (82.3 to 98.5) | 90.9 (72.2 to 97.5) | 44.9 (34.3 to 55.9) | 0.632 (0.565 to 0.704) |
| PV (n=54) | |||||
| AI-PCE | 94.4 (84.9 to 98.1) | 100 (92.3 to 100) | 100 (93 to 100) | 93.9 (83.5 to 97.9) | 0.973 (0.939 to 1.000) |
| CR-PCE | 61.1 (47.8 to 72.9) | 100 (92.3 to 100) | 100 (89.6 to 100) | 68.7 (56.8 to 78.5) | 0.807 (0.740 to 0.873) |
| Colonoscopy | 18.5 (10.4 to 30.8) | 93.9 (85.4 to 97.6) | 71.4 (45.4 to 88.3) | 58.5 (49.0 to 67.4) | 0.562 (0.506 to 0.625) |
| PUE (n=17) | |||||
| AI-PCE | 94.1 (73.0 to 99.0) | 100 (95.6 to 100) | 100 (80.6 to 100) | 98.8 (93.6 to 99.8) | 0.970 (0.900 to 1.000) |
| CR-PCE | 41.2 (21.6 to 64.0) | 100 (95.6 to 100) | 100 (64.6 to 100) | 89.2 (81.3 to 94.1) | 0.705 (0.588 to 0.833) |
| Colonoscopy | 11.8 (3.3 to 34.3) | 98.8 (93.5 to 99.8) | 66.7 (20.8 to 93.9) | 84.5 (76.0 to 90.4) | 0.553 (0.488 to 0.637) |
| PR (n=5) | |||||
| AI-PCE | 100 (56.6 to 100) | 100 (96.1 to 100) | 100 (56.6 to 100) | 100 (96.1 to 100) | 1.000 (1.000 to 1.000) |
| CR-PCE | 80.0 (37.6 to 96.4) | 100 (96.1 to 100) | 100 (51.0 to 100) | 98.9 (94.3 to 99.8) | 0.900 (0.667 to 1.000) |
| Colonoscopy | 60.0 (23.1 to 88.2) | 100 (96.1 to 100) | 100 (43.9 to 100) | 97.9 (92.8 to 99.4) | 0.800 (0.500 to 1.000) |
| P3 (n=7) | |||||
| AI-PCE | 100 (64.6 to 100) | 100 (96.0 to 100) | 100 (64.6 to 100) | 100 (96.0 to 100) | 1.000 (1.000 to 1.000) |
| CR-PCE | 100 (64.6 to 100) | 100 (96.0 to 100) | 100 (64.6 to 100) | 100 (96.0 to 100) | 1.000 (1.000 to 1.000) |
| Colonoscopy | 0 (0 to 35.4) | 100 (96.0 to 100) | – | 93.0 (86.3 to 96.6) | 0.500 (0.500 to 0.500) |
AI-PCE, artificial intelligence-assisted pan-intestinal capsule endoscopy; AUROC, area under the receiver operating characteristic curve; CR-PCE, conventional reading pan-intestinal capsule endoscopy; NPV, negative predictive value; P3, active bleeding; PHL, potentially haemorrhagic lesions; PPV, positive predictive value; PR, haemorrhagic protruding lesions; PUE, ulcers or erosions; PV, vascular lesions.
Table 2. Diagnostic accuracy of AI-PCE versus CR-PCE versus colonoscopy for small bowel and for colon per type of lesion.
| Sensitivity (%) (95% CI) |
Specificity (%) (95% CI) |
PPV (%) (95% CI) |
NPV (%) (95% CI) |
AUROC (%) (95% CI) |
|
|---|---|---|---|---|---|
| PHL | |||||
| Small bowel (n=44) | |||||
| AI-PCE | 95.5 (84.9 to 98.7) | 98.2 (90.6 to 99.7) | 97.7 (87.9 to 99.6) | 96.5 (88.1 to 99.0) | 0.968 (0.929 to 1.000) |
| CR-PCE | 59.1 (44.4 to 72.3) | 100 (93.6 to 100) | 100 (87.1 to 100) | 75.7 (64.8 to 84.0) | 0.797 (0.722 to 0.872) |
| Colon (n=41) | |||||
| AI-PCE | 90.2 (77.5 to 96.1) | 100 (93.9 to 100) | 100 (90.6 to 100) | 93.7 (84.8 to 97.5) | 0.951 (0.900 to 0.989) |
| CR-PCE | 68.3 (53.0 to 80.4) | 100 (93.9 to 100) | 100 (87.9 to 100) | 81.9 (71.5 to 89.1) | 0.843 (0.771 to 0.910) |
| Colonoscopy | 31.7 (19.6 to 46.9) | 88.1 (77.5 to 94.1) | 65.0 (43.3 to 81.9) | 65.0 (54.1 to 74.5) | 0.600 (0.519 to 0.683) |
| PV | |||||
| Small bowel (n=32) | |||||
| AI-PCE | 96.9 (84.3 to 99.5) | 98.5 (92.1 to 99.7) | 96.9 (84.3 to 99.5) | 98.5 (92.1 to 99.7) | 0.977 (0.937 to 1.000) |
| CR-PCE | 56.3 (39.3 to 71.8) | 100 (94.7 to 100) | 100 (84.2 to 100) | 82.9 (73.4 to 89.5) | 0.781 (0.692 to 0.871) |
| Colon (n=33) | |||||
| AI-PCE | 87.9 (72.7 to 95.2) | 98.5 (92.0 to 99.7) | 96.7 (83.3 to 99.4) | 94.3 (86.2 to 97.8) | 0.933 (0.868 to 0.984) |
| CR-PCE | 66.7 (49.6 to 80.2) | 100 (94.6 to 100) | 100 (85.1 to 100) | 85.9 (76.5 to 91.9) | 0.834 (0.750 to 0.914) |
| Colonoscopy | 27.3 (15.1 to 44.2) | 92.5 (83.7 to 96.8) | 64.3 (38.8 to 83.7) | 72.1 (61.8 to 80.5) | 0.600 (0.520 to 0.680) |
| PUE | |||||
| Small bowel (n=12) | |||||
| AI-PCE | 91.7 (64.6 to 98.5) | 100 (95.8 to 100) | 100 (82.4 to 100) | 98.9 (93.9 to 99.8) | 0.957 (0.857 to 1.000) |
| CR-PCE | 50.0 (25.4 to 74.6) | 100 (95.8 to 100) | 100 (60.9 to 100) | 93.6 (86.8 to 97.0) | 0.751 (0.600 to 0.900) |
| Colon (n=8) | |||||
| AI-PCE | 87.5 (52.9 to 97.8) | 100 (96.0 to 100) | 100 (64.6 to 100) | 98.9 (94.2 to 99.8) | 0.938 (0.800 to 1.000) |
| CR-PCE | 37.5 (13.7 to 69.4) | 100 (96.0 to 100) | 100 (43.9 to 100) | 94.8 (88.5 to 97.8) | 0.688 (0.500 to 0.875) |
| Colonoscopy | 12.5 (2.2 to 47.1) | 98.9 (94.1 to 99.8) | 50.0 (9.45 to 90.6) | 92.9 (86.0 to 96.5) | 0.560 (0.480 to 0.700) |
| PR | |||||
| Small bowel (n=1) | |||||
| AI-PCE | 100 (20.7 to 100) | 100 (96.3 to 100) | 100 (20.7 to 100) | 100 (96.3 to 100) | 1.000 (1.000 to 1.000) |
| CR-PCE | 100 (20.7 to 100) | 100 (96.3 to 100) | 100 (20.7 to 100) | 100 (96.3 to 100) | 1.000 (1.000 to 1.000) |
| Colon (n=4) | |||||
| AI-PCE | 100 (51.0 to 100) | 100 (96.2 to 100) | 100 (51.0 to 100) | 100 (96.2 to 100) | 1.000 (1.000 to 1.000) |
| CR-PCE | 75.0 (30.1 to 95.4) | 100 (96.2 to 100) | 100 (43.9 to 100) | 98.9 (94.4 to 99.8) | 0.873 (0.500 to 1.000) |
| Colonoscopy | 75.0 (30.1 to 95.4) | 100 (96.2 to 100) | 100 (43.9 to 100) | 99.0 (94.4 to 99.8) | 0.870 (0.500 to 1.000) |
| P3 | |||||
| Small bowel (n=3) | |||||
| AI-PCE | 66.7 (20.8 to 93.9) | 100 (96.2 to 100) | 100 (34.2 to 100) | 98.9 (94.4 to 99.8) | 0.835 (0.500 to 1.000) |
| CR-PCE | 66.7 (20.8 to 93.9) | 100 (96.2 to 100) | 100 (34.2 to 100) | 98.9 (94.4 to 99.8) | 0.835 (0.500 to 1.000) |
| Colon (n=6) | |||||
| AI-PCE | 100 (60.9 to 100) | 100 (96.1 to 100) | 100 (60.9 to 100) | 100 (96.1 to 100) | 1.000 (1.000 to 1.000) |
| CR-PCE | 100 (60.9 to 100) | 100 (96.1 to 100) | 100 (60.9 to 100) | 100 (96.1 to 100) | 1.000 (1.000 to 1.000) |
| Colonoscopy | 0 (0.0 to 39.0) | 100 (96.1 to 100) | – | 94.0 (87.5 to 97.2) | 0.500 (0.500 to 0.500) |
AI-PCE, artificial intelligence-assisted pan-intestinal capsule endoscopy; AUROC, area under the receiver operating characteristic curve; CR-PCE, conventional reading pan-intestinal capsule endoscopy; NPV, negative predictive value; P3, active bleeding; PHL, potentially haemorrhagic lesions; PPV, positive predictive value; PR, haemorrhagic protruding lesions; PUE, ulcers or erosions; PV, vascular lesions.
In the colon, AI-PCE outperformed CR-PCE in sensitivity (90% vs 68%, p=0.027) and showed a trend towards higher NPV (94% vs 86%, p=0.066), with identical specificity and PPV. When stratified by lesion category, AI-PCE outperformed CR-PCE for vascular lesions and ulcers/erosions, also showing higher sensitivity for haemorrhagic protruding lesions, with similar performance for active bleeding. Compared with colonoscopy, AI-PCE was vastly superior in sensitivity (90% vs 32%, p<0.0001), PPV (100% vs 65%, p<0.001) and NPV (94% vs 65%, p<0.0001).
Figure 4 illustrates some example images of lesions detected by AI-PCE and missed by CR-PCE.
Figure 4. Example images of lesions detected by AI-PCE and missed by CR-PCE. Upper (SB): (a) and (b) angioectasias, (c) ulcer, (d) bleeding. Lower (colon): (e) and (f) angioectasias, (g) ulcer, (h) protruding lesion. AI-PCE, artificial intelligence-assisted pan-intestinal capsule endoscopy; CR-PCE, conventional reading pan-intestinal capsule endoscopy; SB, small bowel.
Discussion
In a cohort of 100 patients with MLGIB, AI-PCE significantly outperformed CR-PCE and colonoscopy for both small bowel and colon PHL detection, especially in sensitivity and NPV, while maintaining excellent specificity. This is the first study to showcase the potential value of AI-enhanced capsule panendoscopy for the detection of PHL, with results indicating higher diagnostic performance compared with both conventional PCE reading and colonoscopy. These findings highlight the clinical relevance of integrating AI into PCE reading. This improvement is particularly important in the context of gastrointestinal bleeding, where missed lesions may have direct therapeutic and prognostic implications.
Therefore, AI-enhanced CE may represent a disruptive change in the approach to MLGIB, becoming a minimally invasive first-line approach for the diagnosis of PHL in both the small bowel and colon.
The advantages of AI may be especially pronounced in PCE due to its unique technical and interpretive challenges. Unlike single-camera small bowel capsules, PCE uses two cameras operating at higher frame rates, producing a larger dataset with twice the visual fields to assess. The examination is longer, covering both small and large bowel, and requires the reader to adapt to the markedly different mucosal appearances between these segments. In addition, capsule movement can be unpredictable, and bowel cleansing is often suboptimal in at least one segment, further increasing reading difficulty. These complexities create an ideal scenario for AI assistance, where convolutional neural networks can efficiently screen large volumes of frames, standardise detection across bowel segments and mitigate reader fatigue.
A recent paper postulated the potential role of PCE as a first-line examination after a negative upper gastrointestinal endoscopy in patients with MLGIB, given its safety profile and accuracy in the detection of both enteric and colonic PHL.2 Indeed, the potential role of PCE in patients with MLGIB has already been demonstrated in previous studies, showing a high NPV and the ability to avoid invasive conventional colonoscopies.21 22 Nevertheless, the results of these studies have never been the subject of a cost-effectiveness analysis. In fact, PCE is a time-consuming examination, with reading times up to 120 min per examination.12 Additionally, there is a risk of missing clinically relevant lesions due to suboptimal bowel preparation, rapid transit time, physician fatigue or even lack of experience in CE reading. Therefore, AI-PCE is a novel alternative for a panenteric assessment of the gastrointestinal tract in patients with MLGIB, given its high diagnostic accuracy while achieving a significant reduction in examination reading times.
Indeed, AI-PCE detected PHL in 60 out of 63 patients, compared with 42 out of 63 in the conventional reading analysis. Moreover, AI-PCE had significantly higher detection rates of vascular lesions, ulcers and erosions compared with CR-PCE. AI-PCE achieved higher NPV than CR-PCE (92% vs 63%), resulting in a low number of missed PHL. Additionally, the increased diagnostic accuracy was achieved with a significant reduction in the examination reading time. Therefore, despite the absence of a cost-effectiveness analysis, AI-PCE appears to be an accurate minimally invasive approach for patients with MLGIB, potentially increasing the value of capsule panendoscopy through more accurate and efficient reading. Although our study did not measure reading times, prior research has shown that AI-assisted CE can significantly reduce review duration.12
Additionally, we compared AI-PCE with conventional colonoscopy in the evaluation of patients with MLGIB. PCE with conventional human reading had been previously shown to be significantly superior to colonoscopy for the detection of PHL in this setting.2 With the use of AI-PCE, these differences are even more remarkable. Indeed, AI-PCE outperformed colonoscopy in the evaluation of colonic PHL, with higher sensitivity (90% vs 32%) and NPV (94% vs 65%). In fact, out of the 41 patients with PHL in the colonic topography, only 13 had them identified with conventional colonoscopy. Interestingly, AI-PCE outperformed colonoscopy in detecting colonic PHL despite the fact that PCE was incomplete in 24% of cases. One possible explanation is that CE allows a longer observation interval during colonic transit compared with the relatively brief inspection during colonoscopy withdrawal. The capsule’s ability to record continuously with two cameras in a liquid-distended colon, permits an antegrade and retrograde caption of images, including behind folds, with an extended viewing time that may facilitate the detection of subtle vascular lesions, small ulcers or intermittent active bleeding that could be missed during a single pass with colonoscopy. Additionally, another consideration must also be made in the comparison between PCE and conventional colonoscopy. Despite being the current gold standard for the evaluation of patients with MLGIB after a negative EGD, colonoscopy is an invasive examination with a non-neglectable risk of adverse events such as bleeding or perforation.23 Also, the frequent use of sedation is associated with both an increased risk of adverse events and higher examination cost. Furthermore, colon CE manufacturing and transportation is responsible for only a small fraction of the carbon footprint associated with diagnostic colonoscopy procedures.24 Therefore, AI-PCE could be a greener alternative to manage these patients, while assuring higher diagnostic accuracy.25
On the other hand, colonoscopy is only able to evaluate the colon and a small segment of the terminal ileum, while PCE is able to explore both the small bowel and colon. In our cohort, 22 out of 100 patients had PHL only in the small bowel, which would be diagnosed only with CE, although this number would be significantly lower recurring only to CR-PCE. Therefore, assuming its higher diagnostic accuracy in the diagnosis of PHL in the colonic topography and the ability to accurately evaluate the small bowel, AI-PCE could be a first-line examination after a negative EGD in patients with MLGIB, with colonoscopy being reserved for patients with colonic PHL that need biopsies or endoscopic treatment. In this cohort, only 41 out of 100 patients could benefit from a colonoscopy for management of colonic PHL lesions.
Nevertheless, there is a need to consider the intrinsic limitations of CE for a panenteric evaluation. First, the colonic evaluation with PCE is dependent on examination completeness. Incomplete examinations, whether due to slow transit or technical malfunction, can reduce the diagnostic yield and necessitate additional investigations, such as conventional colonoscopy or device-assisted enteroscopy. In our cohort, the completion rate for PCE was 76%, meaning that nearly one-quarter of patients required consideration of further procedures. A recent meta-analysis reported a mean completeness rate of 83% for PCE.26 In this context, some patients would require a colonoscopy in a PCE-first approach due to an incomplete colon CE examination. This limitation must be factored into any PCE-first diagnostic strategy. Additionally, PCE evaluation is dependent on good bowel cleansing, such as in the case of colonoscopy. In this context, there is a need for accurate bowel cleansing scales for both the small bowel and colon. In this cohort, the authors used clinically validated scales for cleansing evaluation.20 27 Nevertheless, when considering AI-enhanced CE, specific CNN-based evaluations are required for a trustworthy evaluation of mucosa visualisation and the need to repeat an examination before assuming a negative diagnosis.28 Future studies will integrate both cleansing evaluation and lesion detection in multiple clinical settings. Finally, CE has known limitations in the evaluation of the upper gastrointestinal tract. The absence of insufflation and dependence on peristaltic movements limits the diagnostic accuracy of CE for oesophageal and gastric PHL. In this context, the development of magnetically controlled CE devices could allow for a minimally invasive panenteric evaluation of the gastrointestinal tract, especially if specific CNN models for those topographies are included, allowing for AI-enhanced capsule panendoscopy.29
Some limitations should be acknowledged regarding our study. First, although the dataset was prospectively collected, the analysis was performed retrospectively, on a relatively small sample size. Second, all the patients were from a single tertiary centre, which could be potentially associated with a risk of a demographic bias.30 In order to reduce the impact of this bias, future studies should focus on the prospective clinical validation of AI-PCE in patients from different centres and demographic contexts.30 Third, an AI-PCE first approach is dependent on a complete PCE examination with adequate bowel cleansing. In our cohort, only 76% and 72% of the patients had complete PCE and adequate bowel cleansing for PCE, respectively. Therefore, results must be interpreted in this context and there is a need to increase both the number of complete and adequate PCE examinations before widely implementing AI-PCE in patients with MLGIB. It also should be noted that some lesion subtypes (such as protruding lesions and active bleeding) were infrequent, limiting the statistical power to demonstrate significant differences between AI-PCE and conventional reading. Moreover, while the AI algorithm was externally validated, the study setting involved a single patient cohort, and generalisability to other populations and capsule platforms warrants further evaluation. Finally, cost-effectiveness was not formally assessed in this analysis, and while AI-PCE may reduce reporting time and improve detection, its economic impact requires dedicated study.
In conclusion, AI-PCE was demonstrated, for the first time in full videos, to be superior to CR-PCE and conventional colonoscopy for the detection of both enteric and colonic PHL in MLGIB, assuring higher diagnostic accuracy while offering the potential for reduced examination reading times, as supported by previous literature.12 Taken together, these results support AI-assisted CE as a more accurate and reliable diagnostic tool than either conventional CE reading or colonoscopy in patients with suspected MLGIB. Therefore, AI-PCE may be regarded as a potential new gold standard for PCE reading, as a first-line procedure in MLGIB, offering an opportunity for the evaluation of the small bowel and colon, and reducing the demand for invasive colonoscopies. By combining high accuracy with a minimally invasive and patient-friendly approach, AI-PCE may help streamline diagnostic pathways, reduce the need for multiple procedures and ultimately improve patient outcomes.
Although some limitations remain, particularly regarding incomplete examinations and small numbers in certain lesion categories, our results strongly support the integration of AI into CE workflows. With further validation and cost-effectiveness analyses, AI-PCE has the potential to redefine diagnostic strategies for suspected MLGIB, supporting its integration into routine clinical practice, while offering a reliable, minimally invasive and comprehensive alternative to conventional approaches.
Supplementary material
Footnotes
Funding: The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors.
Provenance and peer review: Not commissioned; externally peer reviewed.
Patient consent for publication: Not applicable.
Ethics approval: This study was conducted according to the Helsinki Declaration and received approval from the Ethics Committee for Health of the coordinating centre, Hospital da Senhora da Oliveira – Guimarães (reference number 187/2024). As this was a secondary analysis of prospectively collected data, and the research was non-interventional with no impact on patient care, the requirement for additional informed consent was waived by the ethics committee.
Data availability statement
Data underlying this article are available from the corresponding author upon reasonable request.
References
- 1.Mascarenhas M, Cardoso H, Macedo G. Artificial Intelligence in Capsule Endoscopy: A Gamechanger for a Groundbreaking Technique. Elsevier; 2023. [Google Scholar]
- 2.Rosa B, Cúrdia Gonçalves T, Moreira MJ, et al. Pan-intestinal capsule endoscopy as first-line procedure in patients with suspected mid or lower gastrointestinal bleeding. Endoscopy. 2024;56:572–80. doi: 10.1055/a-2270-4601. [DOI] [PubMed] [Google Scholar]
- 3.Rosa B, Andrade P, Lopes S, et al. Pan-Enteric Capsule Endoscopy: Current Applications and Future Perspectives. GE Port J Gastroenterol . 2024;31:89–100. doi: 10.1159/000533960. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Patel A, Vedantam D, Poman DS, et al. Obscure Gastrointestinal Bleeding and Capsule Endoscopy: A Win-Win Situation or Not? Cureus. 2022;14 doi: 10.7759/cureus.27137. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Mota J, João Almeida M, Mendes F, et al. A Comprehensive Review of Artificial Intelligence and Colon Capsule Endoscopy: Opportunities and Challenges. Diagnostics (Basel) 2024;14:2072. doi: 10.3390/diagnostics14182072. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Mascarenhas M, Afonso J, Andrade P, et al. Artificial intelligence and capsule endoscopy: unravelling the future. Ann Gastroenterol. 2021;34:300–9. doi: 10.20524/aog.2021.0606. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Pannala R, Krishnan K, Melson J, et al. Artificial intelligence in gastrointestinal endoscopy. VideoGIE. 2020;5:598–613. doi: 10.1016/j.vgie.2020.08.013. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Okagawa Y, Abe S, Yamada M, et al. Artificial Intelligence in Endoscopy. Dig Dis Sci. 2022;67:1553–72. doi: 10.1007/s10620-021-07086-z. [DOI] [PubMed] [Google Scholar]
- 9.Hassan C, Spadaccini M, Iannone A, et al. Performance of artificial intelligence in colonoscopy for adenoma and polyp detection: a systematic review and meta-analysis. Gastrointest Endosc. 2021;93:77–85. doi: 10.1016/j.gie.2020.06.059. [DOI] [PubMed] [Google Scholar]
- 10.Xu H, Tang RSY, Lam TYT, et al. Artificial Intelligence-Assisted Colonoscopy for Colorectal Cancer Screening: A Multicenter Randomized Controlled Trial. Clin Gastroenterol Hepatol. 2023;21:337–46. doi: 10.1016/j.cgh.2022.07.006. [DOI] [PubMed] [Google Scholar]
- 11.Mascarenhas M, Martins M, Afonso J, et al. Deep learning and capsule endoscopy: Automatic multi-brand and multi-device panendoscopic detection of vascular lesions. Endosc Int Open. 2024;12:E570–8. doi: 10.1055/a-2236-7849. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Piccirelli S, Mussetto A, Bellumat A, et al. New Generation Express View: An Artificial Intelligence Software Effectively Reduces Capsule Endoscopy Reading Times. Diagnostics (Basel) 2022;12:1783. doi: 10.3390/diagnostics12081783. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Pennazio M, Rondonotti E, Despott EJ, et al. Small-bowel capsule endoscopy and device-assisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Guideline - Update 2022. Endoscopy. 2023;55:58–95. doi: 10.1055/a-1973-3796. [DOI] [PubMed] [Google Scholar]
- 14.Delvaux M, Friedman S, Keuchel M, et al. Structured terminology for capsule endoscopy: results of retrospective testing and validation in 766 small-bowel investigations. Endoscopy. 2005;37:945–50. doi: 10.1055/s-2005-870266. [DOI] [PubMed] [Google Scholar]
- 15.Mascarenhas M, Mendes F, Ribeiro T, et al. Deep Learning and Minimally Invasive Endoscopy: Panendoscopic Detection of Pleomorphic Lesions. GE Port J Gastroenterol . 2024;31:408–18. doi: 10.1159/000539837. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Mascarenhas M, Ribeiro T, Afonso J, et al. Deep learning and colon capsule endoscopy: automatic detection of blood and colonic mucosal lesions using a convolutional neural network. Endosc Int Open. 2022;10:E171–7. doi: 10.1055/a-1675-1941. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Mascarenhas Saraiva MJ, Afonso J, Ribeiro T, et al. Deep learning and capsule endoscopy: automatic identification and differentiation of small bowel lesions with distinct haemorrhagic potential using a convolutional neural network. BMJ Open Gastroenterol. 2021;8:e000753. doi: 10.1136/bmjgast-2021-000753. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.de Sousa Magalhães R, Arieira C, Boal Carvalho P, et al. Colon Capsule CLEansing Assessment and Report (CC-CLEAR): a new approach for evaluation of the quality of bowel preparation in capsule colonoscopy. Gastrointest Endosc. 2021;93:212–23. doi: 10.1016/j.gie.2020.05.062. [DOI] [PubMed] [Google Scholar]
- 19.Lai EJ, Calderwood AH, Doros G, et al. The Boston bowel preparation scale: a valid and reliable instrument for colonoscopy-oriented research. Gastrointest Endosc. 2009;69:620–5. doi: 10.1016/j.gie.2008.05.057. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Macedo Silva V, Lima Capela T, Freitas M, et al. Small Bowel CLEansing Assessment and Report (SB-CLEAR): Standardizing bowel preparation report in capsule endoscopy. J Gastroenterol Hepatol. 2023;38:747–51. doi: 10.1111/jgh.16086. [DOI] [PubMed] [Google Scholar]
- 21.Mussetto A, Arena R, Fuccio L, et al. A new panenteric capsule endoscopy-based strategy in patients with melena and a negative upper gastrointestinal endoscopy: a prospective feasibility study. Eur J Gastroenterol Hepatol. 2021;33:686–90. doi: 10.1097/MEG.0000000000002114. [DOI] [PubMed] [Google Scholar]
- 22.Carretero C, Prieto de Frías C, Angós R, et al. Pan-enteric capsule for bleeding high-risk patients. Can we limit endoscopies? Rev Esp Enferm Dig. 2021;113:580–4. doi: 10.17235/reed.2020.7196/2020. [DOI] [PubMed] [Google Scholar]
- 23.Levy I, Gralnek IM. Complications of diagnostic colonoscopy, upper endoscopy, and enteroscopy. Best Pract Res Clin Gastroenterol. 2016;30:705–18. doi: 10.1016/j.bpg.2016.09.005. [DOI] [PubMed] [Google Scholar]
- 24.Baddeley R, Arshanapalli S, Turvill J, et al. O20 Quantifying the environmental impact of colonoscopy and colon capsule in the nhs— it’s complicated!. BSG LIVE’24, 17-20 June 2024, ICC Birmingham; 2024. pp. A10–1. [DOI] [Google Scholar]
- 25.Sebastian S, Dhar A, Baddeley R, et al. Green endoscopy: British Society of Gastroenterology (BSG), Joint Accreditation Group (JAG) and Centre for Sustainable Health (CSH) joint consensus on practical measures for environmental sustainability in endoscopy. Gut. 2023;72:12–26. doi: 10.1136/gutjnl-2022-328460. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Rosa B, Donato H, Cúrdia Gonçalves T, et al. What Is the Optimal Bowel Preparation for Capsule Colonoscopy and Pan-intestinal Capsule Endoscopy? A Systematic Review and Meta-Analysis. Dig Dis Sci. 2023;68:4418–31. doi: 10.1007/s10620-023-08133-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.de Sousa Magalhães R, Chálim Rebelo C, Sousa-Pinto B, et al. CC-CLEAR (Colon Capsule Cleansing Assessment and Report): the novel scale to evaluate the clinical impact of bowel preparation in capsule colonoscopy - a multicentric validation study. Scand J Gastroenterol. 2022;57:625–32. doi: 10.1080/00365521.2022.2026463. [DOI] [PubMed] [Google Scholar]
- 28.Mascarenhas Saraiva MJ, Afonso J, Ribeiro T, et al. AI-Driven Colon Cleansing Evaluation in Capsule Endoscopy: A Deep Learning Approach. Diagnostics (Basel) 2023;13:3494. doi: 10.3390/diagnostics13233494. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Jiang B, Pan J, Qian YY, et al. Clinical guideline on magnetically controlled capsule gastroscopy (2021 edition) J Dig Dis. 2023;24:70–84. doi: 10.1111/1751-2980.13173. [DOI] [PubMed] [Google Scholar]
- 30.Franklin G, Stephens R, Piracha M, et al. The Sociodemographic Biases in Machine Learning Algorithms: A Biomedical Informatics Perspective. Life (Basel) 2024;14:652. doi: 10.3390/life14060652. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Data underlying this article are available from the corresponding author upon reasonable request.