ABSTRACT
Objectives
Small bowel capsule endoscopy (CE) produces lengthy videos that are time‐consuming to review and susceptible to missed lesions. We evaluated whether an open‐source, pretrained artificial intelligence (AI) model (SEE‐AI) could improve diagnostic performance and interpretation efficiency compared with conventional reading.
Methods
We retrospectively analyzed 249 PillCam SB3 examinations performed between 2007 and 2022 at six hospitals, using a two‐reader crossover design. SEE‐AI (confidence threshold 0.1) generated annotated videos with bounding boxes for eight lesion categories. The primary endpoints were sensitivity for lesion detection on a per‐lesion and per‐patient basis. Secondary endpoints included specificity, predictive values, overall accuracy, and reading time. A prespecified subgroup analysis evaluated cases of suspected small‐bowel bleeding (SSBB), focusing on Saurin P1+P2 hemorrhagic lesions.
Results
Across 1550 adjudicated lesions, AI‐assisted reading demonstrated higher sensitivity than conventional reading (per‐lesion: 98.8% [1532/1550] vs. 86.4% [1339/1550]; per‐patient: 99.1% [464/468] vs. 80.3% [376/468]; both p < 0.0001). The mean reading time decreased from 17.9 to 13.7 min (p < 0.0001). In SSBB cases (n = 131), sensitivity for P1+P2 lesions improved on both a per‐lesion basis (98.2% [439/447] vs. 82.8% [370/447]) and per‐patient basis (98.6% [145/147] vs. 73.5% [108/147]), with a shorter reading time (14.1 vs. 18.0 min; all p < 0.0001).
Conclusions
In this multicenter evaluation, SEE‐AI significantly improved lesion detection and reduced reading time for CE interpretation, including SSBB cases, while maintaining openness and reproducibility. AI‐assisted reading may reduce clinicians’ workload and support the adoption of SEE‐AI as a practical tool ― and a potential future standard of care ― for small bowel CE.
Trial Registration
N/A.
Keywords: artificial intelligence, capsule endoscopy, gastrointestinal tract, small intestine, suspected small‐bowel bleeding
AI‐assisted reading with the open‐source SEE‐AI model improves lesion detection sensitivity and reduces reading time in small‐bowel capsule endoscopy. In a multicenter retrospective study of 249 cases, SEE‐AI provides visual cues that support physicians during interpretation while preserving physician‐led final decisions.
1. Introduction
Small bowel capsule endoscopy (CE) is a noninvasive modality that enables visualization of the entire small bowel using a swallowed capsule camera [1]. The PillCam SB3 capsule (Medtronic, Minneapolis, MN, USA) measures 11.4 mm × 26.2 mm, weighs 3 g, and acquires 2–6 frames per second depending on small bowel transit [2]. CE is widely used to evaluate small bowel diseases, including suspected small‐bowel bleeding (SSBB), inflammatory bowel disease, and neoplasia [3]. SSBB is the most common indication and frequently originates from small bowel lesions [4]. However, each PillCam SB3 study typically records approximately 8 h of video and around 50,000 images [5], imposing a substantial interpretive burden that may cause reader fatigue and missed lesions [6].
Applications of artificial intelligence (AI) in gastrointestinal endoscopy are expanding, including detection and characterization of neoplasia during upper endoscopy and colonoscopy [7, 8] and assessment of disease activity in ulcerative colitis [9]. Beyond diagnostic performance, AI may also reduce physician workload.
AI is also promising for CE. The RAPID software for PillCam systems includes “TOP100,” which selects 100 frames most likely to contain potential lesions [10, 11]. The Navicam SB system (Ankon, Wuhan, China) includes “ProScan,” which automatically identifies and marks lesions [12]. Although previous studies have reported encouraging results, most were single‐center, used limited datasets, relied on nonpublic algorithms, or focused on specific lesion types. We previously developed SEE‐AI, an object detection model based on YOLOv5 that detects and classifies small bowel lesions [13, 14]. SEE‐AI is publicly available as an open‐source model [15] and overlays bounding boxes with confidence scores (Figure 1 and Video S1). While our previous work demonstrated lesion‐specific sensitivity, whether AI‐assisted reading improves overall diagnostic performance compared with conventional reading has not been evaluated.
FIGURE 1.
Representative capsule endoscopic images annotated by the SEE‐AI system. Each finding (erosion, lymphangiectasis, bleeding, and polyp‐like lesion) is labeled by a bounding box with an assigned confidence score.
Accordingly, using images collected from multiple institutions, we compared AI‐assisted and conventional reading, evaluating lesion detection sensitivity and reading time to clarify the feasibility of clinical implementation.
2. Methods
2.1. Study Design and Participants
We retrospectively analyzed 249 CE examinations performed with the PillCam SB3 system between January 2007 and September 2022 at six institutions: Fukuoka Red Cross Hospital, Hamanomachi Hospital, Kyushu Central Hospital of the Mutual Aid Association of Public School Teachers, Matsuyama Red Cross Hospital, Onga Nakama Medical Association Onga Hospital, and Yamaguchi Red Cross Hospital. The study was approved by the Institutional Review Board of Kyushu University (approval No. 22344‐00). Participation was solicited using an opt‐out approach, and anonymized CE video data and clinical information were retrospectively collected.
2.2. Video Processing and AI Inference
For each case, the segment from duodenal entry to cecal arrival was extracted using RAPID software for PillCam. Videos were exported in MPG format, in which each unique image is stored as five consecutive frames (i.e., four of every five frames are duplicates). To facilitate review and to avoid overcounting redundant frames, the MPG files were converted to MP4 by removing duplicate frames (downsampling). Conversion was performed on macOS 12 (Apple Inc., Cupertino, CA, USA) using Terminal 2.12 and FFmpeg v4.4.1 (FFmpeg Developers; https://ffmpeg.org) with the following command: ffmpeg ‐i input.mpg ‐r 5 output.mp4.
These MP4 files were used for conventional reading. The same videos were uploaded to Google Drive [16] and analyzed on Google Colaboratory (Google Colab), a cloud‐based Python environment with GPU acceleration [17]. SEE‐AI inference used a predefined confidence threshold of 0.1 to prioritize sensitivity, consistent with our previous validation study [14]. In that study, the relationship between confidence rate and model performance was evaluated, and the F1‐score increased rapidly at confidence rates ≤0.1 before showing a more gradual increase. Because SEE‐AI was intended to serve as a reading aid in which physicians review AI‐generated annotations, we selected this low threshold a priori to reduce missed lesions rather than to maximize F1‐score in the present dataset. Based on the AI output, we generated annotated videos with bounding boxes for AI‐assisted reading.
2.3. AI Model and Training
SEE‐AI is an open‐source model trained on anonymized small‐bowel CE videos from 954 patients who underwent PillCam SB3 examinations at Kyushu University Hospital between September 2014 and June 2021. The training dataset comprised 18,481 images across 41 confirmed disease categories, including 12,320 lesion images with 23,033 annotations and 6161 images of normal small‐bowel mucosa (Figure S1 and Table S1) [14].
2.4. Reading Protocol
The 249 cases were divided by case ID into two groups: Group A (IDs 1–125) and Group B (IDs 126–249). Two board‐certified endoscopists of the Japan Gastroenterological Endoscopy Society (Satoshi Miyazono and Takuto Saiki), each with more than 5 years of CE reading experience, participated in a crossover design: reader Satoshi Miyazono performed AI‐assisted reading for Group A and conventional reading for Group B, whereas reader Takuto Saiki performed conventional reading for Group A and AI‐assisted reading for Group B. Within each group, case order was randomized. To minimize potential learning effects, conventional and AI‐assisted reading were alternated so that the same reader did not perform the two methods consecutively. Readers were blinded to clinical information for all readings. Target lesions comprised eight categories: erosion, lymphangiectasis, redness, angioectasia, bleeding, submucosal tumor (SMT), venous lesion, and polyp‐like lesion (Figure 2; Table 1). Representative examples of redness and angioectasia are provided in Figure S2 to facilitate morphological distinction between these two lesion types. When a lesion was identified, readers recorded the finding and total reading time (minutes). For AI‐assisted reading, endoscopists reviewed the entire video with AI‐generated bounding‐box overlays and were not limited to flagged frames. In both methods, videos were manually reviewed without automatic playback.
FIGURE 2.
Representative images of eight types of lesions. (a) Erosion; (b) Lymphangiectasis; (c) Redness; (d) Angioectasia; (e) Bleeding; (f) Submucosal tumor; (g) Venous lesion; (h) Polyp‐like lesion.
TABLE 1.
Definitions of the lesions.
| Classification of lesions | Definition of lesions |
|---|---|
| Erosion | Areas of mucosal damage, such as erosions, ulcers, and notches |
| Lymphangiectasis | Areas containing lymphatic vessels larger than a point |
| Redness | A nonspecific erythematous color change without identifiable vascular structures or typical features of vascular dilatation |
| Angioectasia | A focal erythematous lesion with a visible cluster of dilated superficial vessels, including spider‐like or arborizing vascular patterns |
| Bleeding | Areas of apparent hemorrhage, exclude bile‐colored intestinal fluid |
| Submucosal tumor (SMT) | Areas resembling submucosal tumors |
| Venous lesion | Areas with venous structures |
| Polyp‐like lesion | Elevated lesions with a base or areas of suspected adenoma |
Target lesions comprised eight categories: erosion, lymphangiectasis, redness, angioectasia, bleeding, submucosal tumor (SMT), venous lesion, and polyp‐like lesion.
2.5. Handling of Duplicate Lesions
When the same lesion appeared in multiple frames, identity was determined based on morphology, location, and temporal proximity. Lesions judged identical were counted once, and duplicates were excluded.
2.6. Reference Standard (Adjudication)
Findings recorded concordantly on AI‐assisted and conventional readings were considered “correct.” For discordant results, a third board‐certified endoscopist (Hisamitsu Kaku; >5 years of CE reading experience) adjudicated correctness.
2.7. Definition of False Positives
Diagnostic performance was assessed within an AI‐assisted workflow in which endoscopists retained final responsibility for interpretation. Reader‐level false positives were defined as findings recorded as lesions but not confirmed by the reference standard. AI‐generated boxes rejected by readers were not counted as reader‐level false positives. Separately, all AI‐generated bounding boxes were classified according to whether they corresponded to confirmed lesions; boxes without confirmed lesion correspondence were defined as algorithmic false‐positive boxes.
2.8. Outcomes and Statistical Analysis
The primary endpoints were lesion detection sensitivity on a per‐lesion and per‐patient basis. Secondary endpoints included per‐lesion and per‐patient specificity, positive predictive value, negative predictive value, overall accuracy, and reading time. Sensitivity was compared using McNemar's test, and reading time was compared using the Wilcoxon signed‐rank test. A two‐sided p < 0.05 was considered statistically significant.
Because SSBB is the most frequent indication for CE, we conducted a prespecified subgroup analysis limited to SSBB cases, comparing detection sensitivity and reading time for hemorrhagic lesions. Hemorrhagic lesions were defined according to the Saurin classification as P1 (uncertain hemorrhagic potential) and P2 (high hemorrhagic potential) [18] [Table S2]. The P1 + P2 group included erosion, redness, angioectasia, bleeding, and polyp‐like lesions. All analyses were performed using JMP Pro 17 (SAS Institute Inc., Cary, NC, USA).
3. Results
3.1. Patient Characteristics
Among the 249 cases, the median age was 67 years (interquartile range [IQR], 49–76); 145 patients were male (58.2%), and 104 were female (41.8%). The median number of frames in MP4 files was 10,071 (IQR, 6806–13,931). The indications for CE were SSBB in 131 cases (52.6%), inflammatory bowel disease in 43 (17.3%), and suspected neoplastic lesion in 11 (4.4%) (Table 2). Overall, 213 cases (85.5%) contained at least one of the eight target lesion types, whereas 36 (14.5%) had none. Across all readings, the following lesion counts were recorded (detected by AI‐assisted and/or conventional reading): erosions in 98 patients (825 lesions), lymphangiectasis in 105 (200 lesions), redness in 69 (187 lesions), angioectasia in 57 (92 lesions), bleeding in 31 (74 lesions), SMT in 43 (64 lesions), venous lesions in 38 (55 lesions), and polyp‐like lesions in 27 (53 lesions). The number of cases with each lesion type in Groups A and B is shown in Table S3.
TABLE 2.
Baseline characteristics of patients (n = 249).
| Characteristics | |
|---|---|
| Median age, years (IQR) | 67 (49–76) |
| Male | 145 (58.2%) |
| Median number of video frames in MP4 files (IQR) | 10,071 (6806–13,931) |
| Cases with at least one of the eight types of lesions | 213 (85.5%) |
| Indications for CE | |
| SSBB | 131 (52.6%) |
| Inflammatory bowel disease | 43 (17.3%) |
| ─ Crohn's disease | 38 (15.3%) |
| ─ Behcet's disease | 3 (1.2%) |
| ─ Ulcerative colitis | 1 (0.4%) |
| ─ IBD‐U (unclassified) | 1 (0.4%) |
| Neoplastic lesion | 11 (4.4%) |
| ─ Follicular lymphoma | 4 (1.6%) |
| ─ Metastatic tumor | 2 (0.8%) |
| ─ Diffuse large B‐cell lymphoma | 1(0.4%) |
| ─ Cronkhite–Canada syndrome | 1 (0.4%) |
| ─ Cowden syndrome | 1 (0.4%) |
| ─ Familial adenomatous polyposis | 1 (0.4%) |
| ─ Inflammatory fibroid polyp | 1 (0.4%) |
| Other | 61 (24.5%) |
| Unknown | 3 (1.2%) |
Age, sex, and indication for CE were collected. The number of video frames in the segment from duodenal entry to cecal arrival was measured. SSBB, suspected small‐bowel bleeding; IQR, interquartile range.
3.2. Lesion Detection Sensitivity (Overall Analysis)
Of the 1550 total lesions, 18 (1.16%) were not detected during AI‐assisted reading, whereas 211 (13.6%) were not detected during conventional reading. Per‐lesion sensitivity was 98.8% (1532/1550) for AI‐assisted reading versus 86.4% (1339/1550) for conventional reading (p < 0.0001). This higher sensitivity of AI‐assisted reading was observed across all eight lesion categories (Table 3). Per‐patient sensitivity was 99.1% (464/468) for AI‐assisted reading versus 80.3% (376/468) for conventional reading (p < 0.0001; Table 4). All lesions recorded during AI‐assisted reading had been flagged by SEE‐AI bounding boxes; no lesion was newly identified by the reader alone without an AI flag.
TABLE 3.
Comparison of per‐lesion detection between artificial intelligence (AI)‐assisted and conventional readings in the overall cohort.
| Classification of lesions | AI‐assisted reading | Conventional reading | p‐value * | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| n | Sensitivity | Specificity | PPV | NPV | Accuracy | Sensitivity | Specificity | PPV | NPV | Accuracy | (Sensitivity) | |
| Erosion | 825 | 0.995 | 0.987 | 0.998 | 0.974 | 0.994 | 0.895 | 0.993 | 0.999 | 0.636 | 0.91 | <0.0001 |
| Lymphangiectasis | 200 | 0.98 | 1 | 1 | 0.973 | 0.988 | 0.87 | 1 | 1 | 0.848 | 0.925 | <0.0001 |
| Redness | 187 | 0.973 | 1 | 1 | 0.973 | 0.986 | 0.845 | 1 | 1 | 0.863 | 0.921 | <0.0001 |
| Angioectasia | 92 | 0.967 | 1 | 1 | 0.985 | 0.989 | 0.609 | 1 | 1 | 0.842 | 0.873 | <0.0001 |
| Bleeding | 74 | 0.986 | 1 | 1 | 0.995 | 0.997 | 0.892 | 1 | 1 | 0.965 | 0.973 | 0.0196 |
| SMT | 64 | 1 | 1 | 1 | 1 | 1 | 0.859 | 0.995 | 0.982 | 0.958 | 0.963 | 0.0027 |
| Venous lesion | 55 | 0.982 | 1 | 1 | 0.995 | 0.996 | 0.8 | 1 | 1 | 0.95 | 0.959 | 0.0039 |
| Polyp‐like lesion | 53 | 1 | 1 | 1 | 1 | 1 | 0.906 | 1 | 1 | 0.978 | 0.982 | 0.0253 |
| Total | 1550 | 0.988 | 0.999 | 0.999 | 0.988 | 0.994 | 0.864 | 0.999 | 0.999 | 0.879 | 0.931 | <0.0001 |
NPV, Negative predictive value; PPV, Positive predictive value.
McNemar test. Significant p‐values are indicated in bold.
TABLE 4.
Comparison of per‐patient detection between artificial intelligence (AI)‐assisted and conventional readings in the overall cohort.
| Classification of lesions | AI‐assisted reading | Conventional reading | p‐value * | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| n | Sensitivity | Specificity | PPV | NPV | Accuracy | Sensitivity | Specificity | PPV | NPV | Accuracy | (Sensitivity) | |
| Erosion | 98 | 1 | 0.987 | 0.98 | 1 | 0.992 | 0.765 | 1 | 1 | 0.868 | 0.908 | <0.0001 |
| Lymphangiectasis | 105 | 0.99 | 1 | 1 | 0.993 | 0.996 | 0.829 | 1 | 1 | 0.889 | 0.928 | <0.0001 |
| Redness | 69 | 0.971 | 1 | 1 | 0.989 | 0.992 | 0.812 | 1 | 1 | 0.933 | 0.948 | 0.0045 |
| Angioectasia | 57 | 0.982 | 1 | 1 | 0.995 | 0.996 | 0.649 | 1 | 1 | 0.906 | 0.92 | <0.0001 |
| Bleeding | 31 | 1 | 1 | 1 | 1 | 1 | 0.871 | 1 | 1 | 0.982 | 0.984 | 0.0455 |
| SMT | 43 | 1 | 1 | 1 | 1 | 1 | 0.907 | 1 | 1 | 0.981 | 0.984 | 0.0455 |
| Venous lesion | 38 | 1 | 1 | 1 | 1 | 1 | 0.842 | 1 | 1 | 0.972 | 0.976 | 0.0143 |
| Polyp‐like lesion | 27 | 1 | 1 | 1 | 1 | 1 | 0.852 | 1 | 1 | 0.982 | 0.984 | 0.0455 |
| Total | 468 | 0.991 | 0.999 | 0.996 | 0.997 | 0.997 | 0.803 | 1 | 1 | 0.943 | 0.954 | <0.0001 |
NPV, Negative predictive value; PPV, Positive predictive value.
McNemar test. Significant p‐values are indicated in bold.
3.3. AI‐generated Bounding Boxes and Algorithmic False Positives
Across the 249 analyzed videos, SEE‐AI generated a total of 365,959 bounding boxes, corresponding to a mean of 1469.7 boxes per case. Of these, 134,061 boxes corresponded to actual lesions confirmed by the reference standard, whereas 231,898 boxes did not correspond to confirmed lesions and were therefore classified as algorithmic false‐positive boxes. The lesion‐correspondence rate at the bounding‐box level was 36.6% (134,061/365,959), and the algorithmic false‐positive box rate was 63.4% (231,898/365,959).
3.4. Reading Time
The mean reading time was 13.7 min (range, 3–42) for AI‐assisted reading and 17.9 min (range, 3–67) for conventional reading, which was significantly shorter with AI assistance (p < 0.0001; Figure 3).
FIGURE 3.
Comparison of reading time between artificial intelligence (AI)‐assisted and conventional readings in the overall analysis.
3.5. Subgroup Analysis of SSBB
In the 131 SSBB cases, per‐lesion sensitivity for Saurin P1 + P2 lesions was 98.2% (439/447) for AI‐assisted reading versus 82.8% (370/447) for conventional reading (p < 0.0001; Table 5). Per‐patient sensitivity was likewise higher with AI‐assisted reading: 98.6% (145/147) compared with 73.5% (108/147) for conventional reading (p < 0.0001; Table 6). Mean reading time was 14.1 min (range, 3–40) for AI‐assisted reading and 18.0 min (range, 3–65) for conventional reading (p < 0.0001; Figure 4).
TABLE 5.
Comparison of per‐lesion detection between artificial intelligence (AI)‐assisted and conventional readings in suspected small‐bowel bleeding (SSBB) cases.
| Classification of lesions | AI‐assisted reading | Conventional reading | p‐value * | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| n | Sensitivity | Specificity | PPV | NPV | Accuracy | Sensitivity | Specificity | PPV | NPV | Accuracy | (Sensitivity) | |
| Erosion | 204 | 0.99 | 0.98 | 0.99 | 0.98 | 0.987 | 0.882 | 1 | 1 | 0.797 | 0.919 | <0.0001 |
| Redness | 92 | 0.978 | 1 | 1 | 0.979 | 0.989 | 0.793 | 1 | 1 | 0.833 | 0.898 | 0.0002 |
| Angioectasia | 67 | 0.955 | 1 | 1 | 0.969 | 0.981 | 0.642 | 1 | 1 | 0.795 | 0.85 | <0.0001 |
| Bleeding | 63 | 0.984 | 1 | 1 | 0.991 | 0.994 | 0.889 | 1 | 1 | 0.94 | 0.959 | 0.0339 |
| Polyp‐like lesion | 21 | 1 | 1 | 1 | 1 | 1 | 0.857 | 1 | 1 | 0.976 | 0.979 | 0.0833 |
| Total (P1+P2) | 447 | 0.982 | 0.996 | 0.995 | 0.985 | 0.99 | 0.828 | 1 | 1 | 0.869 | 0.920 | <0.0001 |
NPV, Negative predictive value; PPV, Positive predictive value.
McNemar test. Significant p‐values are indicated in bold.
TABLE 6.
Comparison of per‐patient detection between artificial intelligence (AI)‐assisted and conventional readings in suspected small‐bowel bleeding (SSBB) cases.
| Classification of lesions | AI‐assisted reading | Conventional reading | p‐value * | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| n | Sensitivity | Specificity | PPV | NPV | Accuracy | Sensitivity | Specificity | PPV | NPV | Accuracy | (Sensitivity) | |
| Erosion | 38 | 1 | 0.978 | 0.95 | 1 | 0.985 | 0.658 | 1 | 1 | 0.877 | 0.901 | 0.0003 |
| Redness | 38 | 0.974 | 1 | 1 | 0.989 | 0.992 | 0.763 | 1 | 1 | 0.912 | 0.931 | 0.0114 |
| Angioectasia | 38 | 0.974 | 1 | 1 | 0.989 | 0.992 | 0.711 | 1 | 1 | 0.894 | 0.916 | 0.0039 |
| Bleeding | 22 | 1 | 1 | 1 | 1 | 1 | 0.864 | 1 | 1 | 0.973 | 0.977 | 0.0833 |
| Polyp‐like lesion | 11 | 1 | 1 | 1 | 1 | 1 | 0.727 | 1 | 1 | 0.976 | 0.977 | 0.0833 |
| Total (P1+P2) | 147 | 0.986 | 0.996 | 0.986 | 0.996 | 0.994 | 0.735 | 1 | 1 | 0.929 | 0.940 | <0.0001 |
NPV, Negative predictive value; PPV, Positive predictive value.
McNemar test. Significant p‐values are indicated in bold.
FIGURE 4.
Comparison of reading time between artificial intelligence (AI)‐assisted and conventional readings in suspected small‐bowel bleeding (SSBB) cases.
4. Discussion
In this multicenter study of 249 CE cases, we compared AI‐assisted reading using our open‐source SEE‐AI model with conventional reading. The cohort predominantly consisted of patients undergoing CE for SSBB, and a total of 1550 lesions across eight categories (erosion, lymphangiectasis, redness, angioectasia, bleeding, SMT, venous lesion, and polyp‐like lesion) were analyzed. AI‐assisted reading demonstrated higher sensitivity than conventional reading on both a per‐lesion basis (98.8% vs. 86.4%) and per‐patient basis (99.1% vs. 80.3%), with consistently improved detection across all lesion types. Furthermore, AI‐assisted reading significantly reduced reading time (13.7 vs. 17.9 min; p < 0.0001). In SSBB cases, AI‐assisted reading similarly improved sensitivity for Saurin P1 + P2 lesions (per‐lesion 98.2%, per‐patient 98.6%) while also shortening reading time. Collectively, SEE‐AI appears to enhance both diagnostic performance and efficiency in routine CE interpretation.
Lesion detection sensitivity was consistently higher with AI‐assisted reading in both per‐lesion and per‐patient analyses, in the overall cohort as well as in SSBB cases. Ding et al. reported per‐lesion and per‐patient sensitivities of 99.90% versus 76.89% and 99.88% versus 74.57%, respectively [19], whereas Xie et al. reported per‐lesion sensitivities of 79.3% versus 70.7% [20]. For bleeding‐related indications, Giordano et al. evaluated the TOP100 function of the PillCam SB3 and found that 88.35% of Saurin P2 lesions identified by conventional reading were detected [11]. Spada et al. compared the Navicam ProScan system with conventional reading in suspected small bowel bleeding and reported P1 + P2 lesion sensitivities of 73.7% versus 62.4% [21]. Together, these studies demonstrate that AI‐assisted reading improves lesion detection sensitivity, consistent with the present findings. During CE interpretation, decreased reader concentration due to prolonged review can lead to missed lesions even when visible on the video [6]. The marked improvement in sensitivity with AI‐assisted reading in our study supports the ability of SEE‐AI to reduce missed lesions. Given that SSBB is the principal indication for CE [4], demonstrating a benefit specifically in SSBB highlights the clinical relevance of SEE‐AI. Nevertheless, a small number of lesions were detected only by conventional reading, indicating that AI output should support—not replace—physician judgment.
Although AI‐assisted reading markedly improved sensitivity, a small number of lesions were still missed. In conventional reading, angioectasia showed relatively lower detection, likely due to its small and flat morphology. In contrast, AI assistance improved the detection of such subtle lesions. Lesions missed during AI‐assisted reading were often associated with bubbles or debris, peripheral localization, or partial visualization (Figure S3), which may hinder feature extraction and represent challenging scenarios for AI.
Reading time was significantly reduced overall and in the SSBB subgroup. Ding et al. analyzed 3,280 cases and reported mean reading times of 5.9 ± 2.23 min for AI‐assisted reading versus 96.6 ± 22.53 min for conventional reading [19]. Xie et al. reported similar results (5.4 ± 1.5 vs. 51.4 ± 11.6 min) [20]. In overt bleeding cases, Giordano et al. found reading times of 1.9 versus 23 min with the PillCam SB3 TOP100 function [11], whereas Spada et al. reported 3.8 ± 3.3 versus 33.7 ± 22.9 min using Navicam ProScan [21]. In our study, conventional reading was faster than in most prior reports (17.9 min overall; 18.0 min in SSBB). Although typical reading times for small bowel CE have been reported as 30–90 min [22], Omori et al. demonstrated that reading time depends on both reader experience and capsule generation: beginners required 40.2 ± 10.1 versus 23.7 ± 6.7 min for SB2 versus SB3, whereas experts required 23.2 ± 5.6 versus 11.1 ± 2.9 min [23]. Our readers were experienced endoscopists, and all examinations used SB3, likely contributing to shorter conventional reading times. Conversely, our AI‐assisted reading times were longer than those reported in some previous AI studies, which often limited review to AI‐flagged segments. Because AI‐based lesion detection still has a non‐zero miss rate, clinicians must continue reviewing non‐flagged segments in real‐world practice. In this study, although the entire video was reviewed even during AI‐assisted reading, the bounding‐box overlays provided by the AI served as visual cues that enabled readers to focus their attention more efficiently on lesion‐suspected areas. This ability to prioritize the evaluation of suspicious regions is considered to be a primary factor contributing to the reduction in reading time. Under these pragmatic conditions, AI‐assisted reading significantly reduced reading time, supporting the potential of SEE‐AI to reduce workload in daily clinical practice.
The burden of AI‐generated candidate findings should be considered when interpreting the reduction in reading time. Although SEE‐AI generated many algorithmic false‐positive boxes, AI‐assisted reading still significantly shortened reading time compared with conventional reading. Experienced readers may have rapidly dismissed non‐lesion boxes by considering morphology, temporal continuity, and surrounding frames, while using lesion‐associated boxes as visual cues. However, excessive AI outputs may increase cognitive burden, particularly for less experienced readers. Further model refinement and confidence‐threshold optimization are needed to balance sensitivity with interpretive workload.
A key strength of this study is the use of SEE‐AI, a publicly released, pretrained CE AI model validated using multicenter clinical data [14, 15]. Although many CE AI models have been reported, few are available for independent use or external validation. To our knowledge, this is the first multicenter clinical evaluation of an open‐source, pretrained CE AI model, supporting transparency and reproducibility.
This study has several limitations. First, primary readings were performed by only two Japan Gastroenterological Endoscopy Society–certified endoscopists; therefore, inter‐reader variability—particularly regarding experience level—was not fully assessed. Prior studies have shown differences in detection rates between trainees and experts for erosions and ulcers [24]; future SEE‐AI studies should stratify findings by reader experience. Second, the high number of algorithmic false‐positive boxes remains an important limitation, particularly for less experienced readers. Further studies should evaluate how this burden affects real‐world interpretation efficiency across different levels of reader experience. Third, differentiating redness from angioectasia based solely on capsule findings may be imperfect, despite our definitions requiring visible dilated superficial vessels for angioectasia. Because balloon endoscopic confirmation was not performed, some misclassification cannot be excluded. Fourth, SEE‐AI inference currently requires GPU resources and approximately one hour for video extraction and annotated video generation, creating cost and operational burdens. Fifth, all cases were acquired with PillCam SB3; because image color tone and luminance differ between systems [25], cross‐platform validation is needed.
Future studies should evaluate readers with different experience levels, other capsule platforms, streamlined clinical workflows, and larger datasets with high‐quality annotations.
5. Conclusions
In CE, SEE‐AI improved both diagnostic performance and reading efficiency. AI‐assisted reading has the potential to serve as a new standard of care for the evaluation of small bowel disease.
Author Contributions
Satoshi Miyazono: Conceptualization; Methodology; Data curation; Investigation; Validation; Formal analysis; Visualization; Project administration; Resources; Writing – original draft; Writing – review & editing. Junji Umeno: Conceptualization; Methodology; Data curation; Validation; Formal analysis; Supervision; Writing – review & editing. Tomohiro Nagasue: Conceptualization; Methodology; Data curation; Validation; Formal analysis; Supervision; Writing – review & editing. Takuto Saiki: Data curation; Investigation; Writing – review & editing. Hisamitsu Kaku: Data curation; Investigation; Writing – review & editing. Takehiro Torisu: Conceptualization; Methodology; Data curation; Validation; Formal analysis; Writing – review & editing. Akihito Yokote: Software; Data curation; Writing – review & editing. Keisuke Kawasaki: Data curation; Writing – review & editing. Yutaro Ihara: Data curation; Writing – review & editing. Yuichi Matsuno: Data curation; Writing – review & editing. Noriyuki Imazu: Data curation; Writing – review & editing. Tomohiko Moriyama: Data curation; Writing – review & editing. Ahmed Nashaat Mohamed: Data curation; Writing – review & editing. Katsuya Hirakawa: Data curation; Resources; Writing – review & editing. Hajime Yamagata: Data curation; Resources; Writing – review & editing. Yasuharu Okamoto: Data curation; Resources; Writing – review & editing. Koichi Kurahara: Data curation; Resources; Writing – review & editing. Shinichiro Yada: Data curation; Resources; Writing – review & editing. Akira Harada: Data curation; Resources; Writing – review & editing. Tetsuro Ago: Conceptualization; Methodology; Data curation; Validation; Formal analysis; Supervision; Writing – review & editing. All authors reviewed and approved the final manuscript.
Conflicts of Interest
The authors declare no conflicts of interest.
Funding
The authors have nothing to report.
Ethics Statement
This retrospective study was approved by the Institutional Review Board of Kyushu University (approval No. 22344‐00). Participation was solicited using an opt‐out approach, and anonymized capsule endoscopy video data and clinical information were retrospectively collected.
Consent
Obtained via an opt‐out method.
Supporting information
Supporting Video 1: Sample video demonstrating the detection of erosions.
Supporting Table 1: Disease background of the dataset.
Supporting Table 2: Saurin classification: Categorization of small‐bowel lesions into three groups (P0, P1, and P2) according to their bleeding potential.
Supporting Table 3: Distribution of lesion types in Groups A and B.
Supporting Figure 1: Construction of the SEE‐AI training dataset.
Supporting Figure 2: Representative images of angioectasia and redness. Panels (a–c) show representative examples of angioectasia, whereas panels (d–f) show representative examples of redness.
Supporting Figure 3: Representative examples of lesions missed during AI‐assisted reading. Panels (a–d) show representative examples of lesions missed during AI‐assisted reading. (a) Ulcer with incomplete visualization and surrounding bubbles. (b) Lymphangiectasis with adjacent debris and bubbles. (c) Bleeding with surrounding bubbles and an overall darkened appearance. (d) Venous lesion partially visualized within the frame.
Acknowledgments
The authors have nothing to report.
References
- 1. Ahmed M., “Video Capsule Endoscopy in Gastroenterology,” Gastroenterology Research 15, no. 2 (2022): 47–55, 10.14740/gr1487. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.[cited Sep 23, 2025], Available from: Medtronic . PillCam SB 3 System [Internet] . (2025) https://www.medtronic.com/covidien/en‐nz/products/capsule‐endoscopy/pillcam‐sb‐3‐system.html.
- 3. Hosoe N., Takabayashi K., Ogata H., and Kanai T., “Capsule Endoscopy for Small‐Intestinal Disorders: Current Status,” Digestive Endoscopy 31, no. 5 (2019): 498–507, 10.1111/den.13346. [DOI] [PubMed] [Google Scholar]
- 4. Liao Z., Gao R., Xu C., and Li Z. S., “Indications and Detection, Completion, and Retention Rates of Small‐Bowel Capsule Endoscopy: A Systematic Review,” Gastrointestinal Endoscopy 71, no. 2 (2010): 280–286, 10.1016/j.gie.2009.09.03. [DOI] [PubMed] [Google Scholar]
- 5.Accessed Oct 1, 2025, Available from: Medtronic plc . Medtronic Announces 510(k) Clearance for PillCam Small Bowel 3 @HOME Endoscopy Procedure . Medtronic Newsroom. (2021) https://news.medtronic.com/2021‐11‐15‐Medtronic‐announces‐510‐k‐clearance‐for‐PillCam‐TM‐Small‐Bowel‐3‐HOME‐endoscopy‐procedure.
- 6. Kim S. H., Yang D. H., and Kim J. S., “Current Status of Interpretation of Small Bowel Capsule Endoscopy,” Clinical Endoscopy 51, no. 4 (2018): 329–333, 10.5946/ce.2018.095. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Lee H., Chung J. W., Yun S. C., et al., “Validation of Artificial Intelligence Computer‐Aided Detection on Gastric Neoplasm in Upper Gastrointestinal Endoscopy,” Diagnostics 14, no. 23 (2024): 2706, 10.3390/diagnostics14232706. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Spadaccini M., Menini M., Massimi D., et al., “AI and Polyp Detection During Colonoscopy,” Cancers 17, no. 5 (2025): 797, 10.3390/cancers17050797. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Kim J. E., Choi Y. H., Lee Y. C., et al., “Deep Learning Model for Distinguishing Mayo Endoscopic Subscore 0 and 1 in Patients With Ulcerative Colitis,” Scientific Reports 13, no. 1 (2023): 11351, 10.1038/s41598-023-38206-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Accessed Oct 2, 2025, Available from:Medtronic . PillCam Software v9 — Top 100 Mode . https://www.medtronic.com/covidien/en‐za/products/capsule‐endoscopy/pillcam‐software.html.
- 11. Giordano A., Escapa M., Urpí‐Ferreruela M., et al., “Diagnostic Accuracy of Artificial Intelligence‐Aided Capsule Endoscopy (TOP100) in Overt Small Bowel Bleeding,” Surgical Endoscopy 37, no. 10 (2023): 7658–7666, 10.1007/s00464-023-10273-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Accessed Oct 2025, 2, Available from: AnX Robotics . NaviCam SB Capsule System. (2025) https://www.anxrobotics.com/products/navicam‐sb‐capsule‐system/.
- 13.[cited Sep 23, 2025], Available from: Ultralytics . YOLOv5 [Internet] . (2020) https://github.com/ultralytics/yolov5.
- 14. Yokote A., Umeno J., Kawasaki K., et al., “Small Bowel Capsule Endoscopy Examination and Open Access Database With Artificial Intelligence: The SEE‐Artificial Intelligence Project,” DEN Open 4 (2023): e258, 1, 10.1002/deo2.258. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.[cited Sep 23, 2025], Available from: Google Colaboratory . [Untitled Colab notebook] [Internet] . (2025) https://colab.research.google.com/drive/1mEE5zXq1U9vC01P‐qjxHR2kvxr_3Imz0?usp=sharing.
- 16.[cited Sep 23, 2025], Available from: Google . Google Drive [Internet] . (2025) https://www.google.com/drive/.
- 17.[cited Sep 23, 2025], Available from: Google . Google Colaboratory [Internet] . (2025) https://colab.google/.
- 18. Saurin J. C., Delvaux M., Gaudin J. L., et al., “Diagnostic Value of Endoscopic Capsule in Patients With Obscure Digestive Bleeding: Blinded Comparison With Video Push‐Enteroscopy,” Endoscopy 35, no. 7 (2003): 576–584, 10.1055/s-2003-40244. [DOI] [PubMed] [Google Scholar]
- 19. Ding Z., Shi H., Zhang H., et al., “Gastroenterologist‐Level Identification of Small‐Bowel Diseases and Normal Variants by Capsule Endoscopy Using a Deep‐Learning Model,” Gastroenterology 157, no. 4 (2019): 1044–1054.e5, 10.1053/j.gastro.2019.06.025. [DOI] [PubMed] [Google Scholar]
- 20. Xie X., Xiao Y. F., Zhao X. Y., et al., “Development and Validation of an Artificial Intelligence Model for Small Bowel Capsule Endoscopy Video Review,” JAMA Network Open 5, no. 7 (2022): e2221992, 10.1001/jamanetworkopen.2022.21992. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Spada C., Piccirelli S., Hassan C., et al., “AI‐Assisted Capsule Endoscopy Reading in Suspected Small Bowel Bleeding: A Multicentre Prospective Study,” The Lancet Digital Health 6, no. 5 (2024): e345–e353, 10.1016/S2589-7500(24)00048-7. [DOI] [PubMed] [Google Scholar]
- 22. Akpunonu B., Hummell J., Akpunonu J. D., and Ud Din S., “Capsule Endoscopy in Gastrointestinal Disease: Evaluation, Diagnosis, and Treatment,” Cleveland Clinic Journal of Medicine 89, no. 4 (2022): 200–211, 10.3949/ccjm.89a.20061. [DOI] [PubMed] [Google Scholar]
- 23. Omori T., Hara T., Sakasai S., et al., “Does the PillCam SB3 Capsule Endoscopy System Improve Image Reading Efficiency Irrespective of Experience? A Pilot Study,” Endoscopy International Open 6, no. 6 (2018): E669–E675, 10.1055/a-0599-5852. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Aoki T., Yamada A., Aoyama K., et al., “Clinical Usefulness of a Deep Learning–Based System as the First Screening on Small‐Bowel Capsule Endoscopy Reading,” Digestive Endoscopy 32, no. 4 (2020): 585–591, 10.1111/den.13517. [DOI] [PubMed] [Google Scholar]
- 25. Van de Bruaene C., De Looze D., and Hindryckx P., “Small Bowel Capsule Endoscopy: Where Are We After Almost 15 Years of Use?” World Journal of Gastrointestinal Endoscopy 7, no. 1 (2015): 13–36, 10.4253/wjge.v7.i1.13. [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
Supporting Video 1: Sample video demonstrating the detection of erosions.
Supporting Table 1: Disease background of the dataset.
Supporting Table 2: Saurin classification: Categorization of small‐bowel lesions into three groups (P0, P1, and P2) according to their bleeding potential.
Supporting Table 3: Distribution of lesion types in Groups A and B.
Supporting Figure 1: Construction of the SEE‐AI training dataset.
Supporting Figure 2: Representative images of angioectasia and redness. Panels (a–c) show representative examples of angioectasia, whereas panels (d–f) show representative examples of redness.
Supporting Figure 3: Representative examples of lesions missed during AI‐assisted reading. Panels (a–d) show representative examples of lesions missed during AI‐assisted reading. (a) Ulcer with incomplete visualization and surrounding bubbles. (b) Lymphangiectasis with adjacent debris and bubbles. (c) Bleeding with surrounding bubbles and an overall darkened appearance. (d) Venous lesion partially visualized within the frame.