Abstract
Objectives:
False positive detection (FPD) is a clinically relevant challenge that must be solved to ensure the safe and effective integration of artificial intelligence (AI) into colonoscopy. This study aimed to develop a novel AI system for quantifying FPD and to compare FPD rates between two commercially available AI platforms.
Methods:
From October 2022 to May 2025, a total of 436,637 images were collected from 78 patient colonoscopy videos: 38 videos (340,121 images) using CAD EYE (Fujifilm, Tokyo, Japan) and 40 videos (96,516 images) using EndoBRAIN (Olympus, Tokyo, Japan). Using these images, YOLOv8 convolutional neural networks were applied to construct original AI model for FPD analysis: ARBS-YOLOv8 (with active learning). The performance of the model was validated using the CAD EYE and EndoBRAIN clinical videos.
Results:
For CAD EYE validation, ARBS-YOLOv8 achieved an accuracy, sensitivity, specificity, precision, and F1 score of 99.22%, 99.51%, 99.97%, 99.96%, and 99.90%, respectively. For EndoBRAIN validation, the ARBS-YOLOv8 results were 99.97%, 99.84%, 99.97%, 99.97%, and 99.90%, respectively. The false-positive and false-negative rates of FPD detection were only 0.04% and 0.03% for CAD EYE and 0.03% and 0.03% for EndoBRAIN. In 13 collected clinical videos (7 CAD EYE, 6 EndoBRAIN), ARBS-YOLOv8 identified total FPD rates of 8.08% for CAD EYE and 2.47% for EndoBRAIN, showing a significant difference (p<0.001).
Conclusion:
We developed the first AI system capable of objectively measuring FPD during colonoscopy under two commercially available AI platforms. The model validated performance differences between CAD EYE and EndoBRAIN, providing a basis for FPD reduction and safe AI endoscopy.
Keywords: artificial intelligence, false positive detection, colonoscopy, CADe
Introduction
Colorectal cancer remains one of the leading causes of cancer incidence and mortality worldwide, and colonoscopy plays a central role in its prevention and early detection[1]. The endoscopic identification and resection of adenomas are directly linked to the prevention of colorectal cancer development, and numerous studies have established the adenoma detection rate (ADR) as a key quality indicator that strongly correlates with patient outcomes[2]. In recent years, the rapid introduction of artificial intelligence (AI)-based computer-aided detection (CADe) systems has markedly improved ADR[3-6]. Recent guidelines from the American Gastroenterological Association (AGA) and a position statement by the European Society of Gastrointestinal Endoscopy (ESGE) support a moderate impact of AI on ADR[7,8]. However, this benefit is accompanied by a new challenge: an increase in false positive detection (FPD). Excessive FPDs can lead to unnecessary endoscopic maneuvers, prolonged withdrawal times, increased costs, and greater patient discomfort.
Several previous reports have examined factors contributing to FPD[9,10]. These include bowel wall artifacts, such as mucosal folds, normal mucosa, diverticula, as well as bowel contents, such as stool, bubbles, and bleeding. While earlier studies focused primarily on improving sensitivity and ADR, recent attention has shifted toward a more balanced performance evaluation, incorporating specificity and positive predictive value (PPV). Randomized controlled trials from both Europe and Japan have consistently demonstrated the effectiveness of AI-assisted colonoscopy in improving ADR and highlighted a measurable rise in false positive alerts[11,12].
FPD is a clinically relevant challenge that must be addressed to ensure the safe and effective integration of AI into routine endoscopic practice. FPD is an important indicator for assessing the quality of AI systems. However, the calculation of FPD is performed manually by endoscopists, and definitions of FPD are different in each study. The present study aims to develop an original AI for evaluating FPD in two commercially available platforms, CAD EYE (Fujifilm Co., Tokyo, Japan) and EndoBRAIN (Olympus Co., Tokyo, Japan).
Methods
From October 2022 to May 2025, a total of 436,637 images were collected from 78 patient colonoscopy videos, with complete colonoscopic observation using the AI systems: CAD EYE (Fujifilm Co., Tokyo, Japan) for 38 patients (340,121 images) and EndoBRAIN (Olympus Co., Tokyo, Japan) for 40 patients (96,516 images). These data were used to develop a novel AI model for evaluating FPD (Figure 1). The development process of an original AI consisted of four datasets: a basic training set, an active learning training set, a validation set, and a test set. The basic training sets included CAD EYE: 16 patient videos (1,893 images); EndoBRAIN: 23 patient videos (997 images) (Table 1). The active learning training sets included CAD EYE: 5 patient videos (234 images); EndoBRAIN: 5 patient videos (235 images). The validation sets included CAD EYE: 10 patient videos (16,345 images with annotation boxes and 230,896 images without annotation boxes); EndoBRAIN: 5 patient videos (6,140 images with annotation boxes, and 28,608 images without annotation boxes). The test sets included CAD EYE: 7 patient videos, 90,753 images; EndoBRAIN: 7 patient videos, 60,536 images. Only videos without polyps were selected for the test sets.
Figure 1.
Flowchart of study.
Table 1.
The Data Sets of Developing Original AI.
| Overall | Basic training sets | Active learning training sets | Validation sets | Test sets | |
|---|---|---|---|---|---|
| Total | 78 patients videos 436,637 images |
39 videos 2,890 images |
10 videos 469 images |
15 videos 281,989 images |
14 videos 151,289 images |
| CAD EYE | 38 patients video 340,121 images |
16 videos 1,893 images |
5 videos 234 images |
10 videos 16,345 images with AB 230,896 images without AB |
7 videos 90,753 images without polyps |
| EndoBRAIN | 40 patients video 96,516 images |
23 videos 997 images |
5 videos 235 images |
5 videos 6,140 images with AB 28,608 images without AB |
7 videos 60,536 images without polyps |
AB: annotation box
We selected YOLOv8 as convolutional neural networks (CNNs) due to its optimal balance between accuracy and speed. YOLOv8 is an innovative object detection algorithm proposed by Ultralytics et al. in 2023[13]. This algorithm comprises of five models, each with varying depths and widths. Compared with the previous YOLO model, YOLOv8 incorporates several improvements to enhance model performance. A customized AI model such as RBS-YOLOv8 (without active learning) and ARBS-YOLOv8 (with active learning) were specifically designed for analyzing CAD EYE and EndoBRAIN videos[14].
Network Architecture Optimization
In the architecture of the proposed AI model, ARBS-YOLOv8, input frames are processed through RepVGGBlock and C2F modules with an SPPF layer, and features are then fused in the Neck before passing to the Head[13,14] (Figure 2a). To build an efficient and accurate detection model, we introduced two key modifications to the standard YOLOv8 architecture[13] and developed ARBS-YOLOv8[14]. First, in the backbone network, we replaced the original Stem layer and the first standard convolutional module (Conv) of each stage with the RepVGGBlock[15] (Figure 2b). Using a unique structural re-parameterization design, RepVGGBlock achieved a lightweight architecture that balances computational efficiency with network depth, while also improving the capacity for feature representation[15]. Second, we substituted the conventional CIoU loss[13] with the SIoU loss function[16] to optimize bounding-box regression. Unlike CIoU, SIoU introduces directional constraints and an angle penalty term in addition to overlap and scale considerations. This design accelerated the alignment between predicted and ground-truth boxes, leading to faster convergence and improved localization precision. The detailed definitions of the distance and shape costs used in SIoU are provided according to a previous report[16].
Figure 2.
Structure of our original AI, ARBS-YOLOv8 with RepVGGBlock.
a. Structure of ARBS-YOLOv8, b. Structure of RepVGGBlock.
Active Learning-Based Training and Optimization
To further enhance generalization in complex scenarios and reduce false detections, we designed an iterative training strategy based on active learning. Active learning is an advanced deep learning strategy that selectively incorporates the most informative samples, typically those with the highest prediction uncertainty into the training set, thereby maximizing model performance under limited annotation costs. In this study, we used this strategy to identify and exploit false positive samples.
The flowchart of the active learning strategy is described below and its implementation consists of the following steps:
•Baseline model training: The RBS-YOLOv8, after architectural optimization, was initially trained using a standard training and validation set to establish a baseline model.
•Uncertainty sample selection: The baseline model was applied to an independent, unlabeled dataset to identify samples with high uncertainty, where the model is prone to confusion.
•Expert review and annotation: Gastroenterology experts manually reviewed the predictions, focusing on verifying all false positive cases.
•Hard negative augmentation: Confirmed false positive samples were incorporated into the dataset as hard negatives with a dedicated “false positive” label, thereby enhancing the model's ability to discriminate between true targets and visually similar non-target regions.
•Iterative retraining: The augmented dataset was used to retrain the model, enabling it to learn from previously misclassified false positives and improving its discrimination capacity.
•Performance evaluation: The iteratively refined model was comprehensively evaluated on an independent test set and compared against the baseline model to quantitatively assess the effectiveness of the active learning strategy.
Through this iterative process, the RBS-YOLOv8 model was optimized into the ARBS-YOLOv8, which represents the final model adopted in this study.
Device and Hyperparameters
This study was conducted on a workstation with an IntelⓇ Core™ i7-13700F processor, a NVIDIAⓇ GeForce™ RTX 4070 GPU with 12 GB memory, and 32 GB of RAM. The confidence threshold for evaluation metrics was set as 0.4 according to our pilot study[14]. The training was carried out using 100 epochs, a batch size of 16, and an image resolution of 640 × 640. Automatic Mixed Precision (AMP) was employed for training all models.
Outcomes
The accuracy, precision, sensitivity, specificity, and F1 score were calculated to evaluate the models' performance in detecting annotation boxes. Precision was defined as true positives (TP) / (TP + false positives (FP)) and the F1 was calculated as 2×precision×sensitivity / (precisions + sensitivity). The false positive rate (FPR) and false negative rate (FNR) were also calculated as additional indicators of detection accuracy. Additionally, complete colonoscopic observation videos from both CAD EYE and EndoBRAIN systems that contained no polyps were examined to calculate the FPD rate. Additionally, the contents of false-positive detections were analyzed using eligible colonoscopy videos with good bowel preparation from CAD EYE and EndoBRAIN after fine-tuning our original AI. False-positive contents were categorized as bowel wall-related findings (folds, ileocecal valve, appendiceal orifice, normal mucosa, diverticula, suction, scar, erosions, blood vessels, hemorrhoid bleeding, erosions, postoperative anastomosis, ulcer) and bowel contents (stool, mucus, bubbles, water, light reflections). This analysis was performed by three expert endoscopists (>2000 screening colonoscopies each) in accordance with a previous report, up to a maximum of 200 FPD events per system[10]. The final classification was determined through discussion until consensus was reached.
To eliminate the potential influence of differences in frame rates between CAD EYE and EndoBRAIN videos, all videos were pre-normalized to 30 frames per second (fps) prior to analysis. The AI model used in this study was specifically developed to evaluate false-positive detections (FPDs) of existing CADe systems and was not intended to function as a real-time diagnostic system. Accordingly, system-specific delays between lesion recognition and visualization of bounding boxes were not considered relevant to the objectives of this analysis. All videos were analyzed offline using a frame-by-frame batch-processing approach. Bounding boxes were generated exclusively for frames in which detection outputs were present, and all frames were subsequently recombined to reconstruct the final videos. Under these standardized conditions, differences in frame rate and system-related display delays did not influence frame-based FPD metrics or FPD per minute.
This study was approved by the Ethics Committee of Kyoto Prefectural University of Medicine (ERB-C-1704-5, Approval date: June 5, 2024) and conducted in accordance with the principles of the World Medical Association Declaration of Helsinki. This was a retrospective study and patients' informed consent was obtained with the option opt-out from participating. Informed consent for colonoscopic procedures with AI was obtained from all participants.
Results
According to the study flow of making the original AI, validation was performed for the two original AIs, RBS-YOLOv8 (without active learning) and ARBS-YOLOv8 (with active learning) (Figure 1, 2, Table 2). For the CAD EYE validation set, the accuracy, sensitivity, specificity, precision, and F1 score of ARBS-YOLOv8 were 99.22%, 99.51%, 99.97%, 99.96%, and 99.90%, showing an improved performance compared with RBS-YOLOv8. For the EndoBRAIN validation set, the corresponding values for ARBS-YOLOv8 were 99.97%, 99.84%, 99.97%, 99.97%, 99.90%. The FPR and FNR of ARBS-YOLOv8[14] were 0.04% and 0.03% for CAD EYE and 0.03% and 0.03% for EndoBRAIN respectively (Figure 3).
Table 2.
Validation of the Original AI, RBS-YOLOv8 and ARBS-YOLOv8 for Examining Annotation Box with CAD EYE and EndoBRAIN.
| Images | AI model | Accuracy | Sensitivity | Specificity | Precision | F1 Score | FPR | FNR |
|---|---|---|---|---|---|---|---|---|
| CAD EYE | RBS-YOLOv8 | 99.22% | 89.43% | 99.98% | 99.16% | 94.41% | 0.84% | 0.02% |
| ARBS-YOLOv8 | 99.22% | 99.51% | 99.97% | 99.96% | 99.90% | 0.04% | 0.03% | |
| EndoBRAIN | RBS-YOLOv8 | 99.78% | 98.79% | 99.97% | 99.74% | 99.38% | 0.26% | 0.03% |
| ARBS-YOLOv8 | 99.97% | 99.84% | 99.97% | 99.97% | 99.90% | 0.03% | 0.03% |
FPR: false positive rate, FNR: false negative rate
Figure 3.
True detection and false detection with our original AI, ARBS-YOLOv8 for false positive detection of CAD EYE and EndoBRAIN.
a. True detection examined with ARBS-YOLOv8 for CAD EYE. b. False detection for CAD EYE. c. True detection for EndoBRAIN. d. False detection for FPD of EndoBRAIN.
Among 14 videos of clinical cases used for correction, 7 CAD EYE and 6 EndoBRAIN videos were examined with the original AI and ARBS-YOLOv8 (Table 3, Video). One EndoBRAIN video was damaged and could not be evaluated. The FPD rates were 8.08±4.09% for CAD EYE and 2.47±1.16% (p=0.004) for EndoBRAIN. The FPD/min (mean±SD) were 125.9 ± 75.2 and 44.5 ± 21.1 (p=0.013), respectively.
Table 3.
False Positive Detection of Clinical Cases with CAD EYE and EndoBRAIN Examined by the Original AI.
| Case number | Time length | Total Frames | ABs Counts/colonoscopy | ABs Frames/colonoscopy | FPD (Percentage of frames detected with ARBS-YOLOv8) | FPD/minutes |
|---|---|---|---|---|---|---|
| CAD EYE1 | 5:45 | 10361 | 427 | 388 | 3.74% | 67.5 |
| CAD EYE2 | 8:00 | 4497 | 602 | 498 | 11.07% | 62.3 |
| CAD EYE3 | 5:42 | 10237 | 833 | 590 | 5.76% | 103.5 |
| CAD EYE4 | 8:00 | 14428 | 2224 | 1608 | 11.14% | 201.0 |
| CAD EYE5 | 7:58 | 14370 | 1330 | 1178 | 8.20% | 148.0 |
| CAD EYE6 | 8:23 | 15086 | 2947 | 2066 | 13.69% | 246.5 |
| CAD EYE7 | 7:27 | 13400 | 414 | 394 | 2.94% | 52.9 |
| Total (mean±SD) | 7.3±1.1 | 1253±983 | 960±669 | 8.08±4.09%* | 125.9±75.2*** | |
| EndoBRAIN1 | 2:15 | 4053 | 102 | 87 | 2.15% | 38.7 |
| EndoBRAIN2 | 4:57 | 8920 | 79 | 76 | 0.85% | 15.4 |
| EndoBRAIN3 | 5:31 | 9901 | 221 | 218 | 2.20% | 39.6 |
| EndoBRAIN4 | 4:40 | 8400 | 333 | 317 | 3.77% | 68.0 |
| EndoBRAIN5 | 2:58 | 5385 | 217 | 210 | 3.90% | 70.9 |
| EndoBRAIN6 | 8:27 | 15160 | 306 | 295 | 1.95% | 34.9 |
| Total (mean±SD) | 4.8±2.1 | 209±103 | 200±101 | 2.47±1.16%** | 44.5±21.1**** |
AB: annotation box, FPD: false positive detection
*vs. **: p=0.004, ***vs. *****: p=0.013
The three most frequent causes of false-positive detections differed between the two systems (Table 4). In CAD EYE, the most frequent FPD contents were folds (39.4%), normal mucosa (16.3%), and blood vessels (6.8%) (Figure 4a, 4b, 4c). In contrast, in EndoBRAIN, the most frequent contents were folds (33.9%), suction (19.2%), and normal mucosa (15.2%) (Figure 4d, 4e, 4f). In both systems, folds were the most common FPD content.
Table 4.
The Contents of False Positive Detection of Clinical Cases with CAD EYE and EndoBRAIN Examined by an Experts Endoscopist.
| The contents of FPD | CAD EYE | EndoBRAIN |
|---|---|---|
| folds, n (%) | 121 (39.4) | 76 (33.9) |
| normal mucosa, n (%) | 50 (16.3) | 34 (15.2) |
| blood vessel, n (%) | 21 (6.8) | 2 (0.9) |
| stool, n (%) | 19 (6.2) | 10 (4.5) |
| light reflections, n (%) | 17 (5.5) | 5 (2.2) |
| ileocecal valve, n (%) | 13 (4.2) | 6 (2.7) |
| suction, n (%) | 13 (4.2) | 43 (19.2) |
| erosions, n (%) | 11 (3.6) | 5 (2.2) |
| mucus, n (%) | 7 (2.3) | 7 (3.1) |
| hemorrhoids, n (%) | 6 (2.0) | 17 (7.6) |
| water, n (%) | 6 (2.0) | 0 (0) |
| bubbles, n (%) | 5 (1.6) | 1 (0.4) |
| ulcer, n (%) | 5 (1.6) | 6 (2.7) |
| bleeding, n (%) | 4 (1.3) | 4 (1.8) |
| scar, n (%) | 4 (1.3) | 1 (0.4) |
| diverticula, n (%) | 3 (1.0) | 2 (0.9) |
| appendiceal orifice, n (%) | 2 (0.7) | 0 (0) |
| postoperative anastomosis, n (%) | 0 (0) | 5 (2.2) |
| TOTAL | 307 | 224 |
| Procedure time (mins) | 36.0 | 61.6 |
| FPD/min | 8.5 | 3.6 |
FPD: false positive detection
Figure 4.
The three most frequent contents of false positive detection in CAD EYE and EndoBRAIN.
a. CAD EYE. Folds. b. CAD EYE. Normal mucosa. c. CAD EYE. Blood vessel. d. EndoBRAIN. Folds. e. EndoBRAIN. Suction. f. EndoBRAIN. Normal mucosa.
Discussion
In the present study, we developed a high-performance AI algorithm capable of objectively and automatically quantifying FPD during colonoscopy. Our validation process demonstrated an extremely high accuracy in reproducing the annotation boxes of two widely implemented commercial CADe systems, CAD EYE and EndoBRAIN. When applied to real-world colonoscopy videos, the AI revealed that the FPD of CAD EYE and EndoBRAIN (mean±SD) were 8.08±4.09 and 2.47±1.16, respectively (p=0.004). To our knowledge, this is the first study to establish a reproducible and automated methodology for measuring FPD across CADe systems in a controlled manner.
The clinical implications of our findings are multifold. First, although CADe systems consistently improve ADR, their widespread adoption necessitates careful consideration of the cognitive burden they may impose. FPs can arise from residual fecal debris, bubbles, mucosal folds, or light reflections. Frequent false alarms may reduce endoscopists' trust in AI output, leading to “alert fatigue,” a phenomenon well-described in other clinical decision support contexts[17]. Second, excessive FPD could potentially prolong withdrawal time, though several reports suggest that AI-assisted colonoscopy does not significantly increase withdrawal duration in controlled trials[5,18]. In high-volume clinical practice, however, the cumulative effect of frequent FPs could influence procedure efficiency. Third, FPD may influence trainee education. For novice endoscopists, repeated FPD may impair learning by reinforcing incorrect visual cues. Our study demonstrated measurable differences between commercially available systems, raising the possibility that certain platforms may achieve a more favorable balance between sensitivity and specificity. Accordingly, the development of an AI algorithm designed to objectively quantify FPD during colonoscopy, as presented in our study, is clinically important for optimizing CADe performance.
FPD in colonoscopy has been evaluated using various methodologies, and reported values vary considerably depending on the measurement approach[19,20] (Table 5). Frame-based analyses provide highly objective and reproducible metrics, with reported false positive per frame ranging from 0.12% to 0.16%[21,22]. However, this method requires labor-intensive frame-by-frame annotation, which limits its feasibility in large-scale studies. Lesion- or cluster-based evaluations, in which consecutive detections of the same lesion are counted as a single false positive, yield lower values of about 0.3-0.46 FP per colonoscopy[23,24]. These are considered more clinically meaningful, although the results may fluctuate depending on the definition of a cluster. Time-based assessments, typically expressed as FP per minute, have reported values of 0.16-0.2 FP/min, corresponding to approximately one false alarm every 5-6 minutes of withdrawal time[25,26]. This approach is intuitive and reflects the procedural disturbance for endoscopists, but it remains influenced by inspection time and withdrawal speed. Finally, case-based evaluations, in which endoscopists record the number of false detections during an examination, report substantially higher values, averaging 4.2-25.7 FP per colonoscopy[27-30]. In the present study, we calculated both FP rate per frame and FP number per colonoscopy by using our AI to detect all frames containing annotation boxes. As a result, our FPD values were higher than those reported in previous studies. While prolonged FPD may occasionally disrupt endoscopists' observations, its influence is inherently subjective; thus, we aimed to evaluate it objectively by calculating all FPDs.
Table 5.
Literature Review of Reports about False Positive Detection of AI in Colonoscopy.
| Kinds of FPD | Representative report Author | Reported values | Merit | Demerit |
|---|---|---|---|---|
| FP rate per frame | Misawa, 2018 Gong, 2020 Ours |
0.12 0.16 2.47-8.08 |
Highly objective, reproducible; suitable for head-to-head AI comparisons | Extremely labor-intensive (requires full-frame annotation); less reflective of clinical experience |
| FP cluster analysis | Wang, 2000 Hassan, 2020 |
0.3 0.46 |
Clinically more relevant; avoids overcounting consecutive detections of the same lesion | Dependent on cluster definition (distance, frame interval); results may vary by criteria |
| FP number per colonoscopy per minute | Byrne, 2019 Su, 2020 |
0.2 0.16 |
Intuitive for endoscopists; reflects procedural disturbance | Influenced by inspection time and withdrawal speed; requires time normalization |
| FP number per colonoscopy | Repici, 2020 Yamada, 2022 Spadaccini, 2022 Okumura, 2024 Ours |
26 20.5 25.5-25.7 4.2-18.1 200-960 |
Closest to real-world endoscopist perception; directly reflects clinical burden | Subjective, operator-dependent; large interobserver variability |
FPD: false positive detection, FP: false positive
Several studies have investigated the underlying cause for FPD during colonoscopy. The most common sources are normal anatomic structures, such as the appendiceal orifice, ileocecal valve, and rectal folds, which are frequently misinterpreted as polyps[17,18,30]. Residual stool, bubbles, and fluid are also major contributors, accounting for up to 20-30% of false alarms in some series[19,21]. We analyzed the contents of FPDs and found that folds were the most frequent category for both CAD EYE and EndoBRAIN. However, the distribution of FPD contents differed slightly between the two AI systems. Additionally, our previous study demonstrated that severe bubbles occurred more frequently in the proximal colon (16%), more than initially expected[31]. Optical artifacts, including light reflection, halation, vascular shadows, and motion blur, have likewise been reported as important causes of FPD[22,23]. In addition, limitations in training datasets―particularly inadequate representation of poor bowel preparation or background mucosa resembling hyperplastic polyps―can further increase the likelihood of false positives[10,24]. Collectively, these findings suggest that most FPDs arise from a combination of anatomic and optical factors, and that optimization of both image acquisition and training data is essential to mitigate this problem. In the current study, we did not analyze the reasons for FPD. However, understanding these causes remains crucial for endoscopists applying AI in routine clinical practice.
Notably, recent evidence suggests a post-upgrade version of the same AI could decrease number of FPs per colonoscopy significantly than a pre-upgrade version (4.2±3.7 vs. 18.1±11.1, p<0.001)[26]. We have previously published several studies on CAD EYE and other original AI models for colonoscopy, providing detailed insights into FPD[32-35]. For example, using an original Fujifilm-based AI, the FPD rate under white-light imaging was 7.8% at an AI confidence threshold of ≥10% and 2.8% at ≥30%[32]. Further, we demonstrated that FPD rates could be improved by refining the training dataset to include defocused polyp images[33]. These findings indicate that FPD can be reduced through optimization of confidence thresholds, training data content, and algorithm design.
Although the potential applications of the proposed framework―such as regulatory evaluation, metric standardization, and AI vigilance―remain conceptual at present, several technical and clinical steps are required to translate this methodology into practical implementation[36]. First, multi-center validation using heterogeneous endoscopic datasets is essential to confirm generalizability. Second, standardized definitions and taxonomies of false-positive patterns need to be established to ensure reproducibility and comparability across studies. Third, the robustness of the framework should be assessed across different CADe architectures and imaging modalities. Finally, integration of this methodology into quality control and post-market surveillance workflows will be necessary for real-world deployment. Addressing these steps may help bridge the gap between methodological development and future clinical and regulatory applications. An automated FPD quantification platform provides an ideal instrument for such monitoring. Additionally, technological advances in endoscopic systems with improvement in brightness and reduced halation[37]- may also help reduce FPD rates. Future studies should explore how these hardware and software improvements interact to further minimize false detections.
In an era where numerous AI systems are rapidly emerging, objective comparison and ranking of AI performance are increasingly important. Key evaluation domains include true polyp detection, FPD, generalizability, usability, and cost. The AI developed in this study is specifically designed to objectively evaluate FPD. While previous studies have assessed false-positive alerts, FPD has often been defined subjectively based on the appearance of annotation boxes[10,29,30]. In contrast, our AI enables objective quantification by counting annotation boxes and classifying them according to their duration. Although this approach is currently limited to polyp-free videos, false-positive alerts are most problematic in the absence of true lesions. Frequent FPDs may prolong observation time and cause endoscopist fatigue, reducing clinical utility even when true-detection sensitivity is high[30]. Our study demonstrated that, even in a limited sample, the frequency and patterns of FPDs differed between CAD EYE and EndoBRAIN, indicating that such differences are measurable and system dependent. These differences may influence endoscopist performance and potentially downstream clinical outcomes. Excessive FPDs may also hinder training, particularly for non-expert endoscopists. Therefore, while this framework is not intended for direct clinical use, it provides a useful and objective tool for developers and researchers to support AI development, comparison, quality control, quality assurance, and post-market surveillance in real-world practice. Additionally, the extremely high F1 score (99.0%) for CAD EYE and EndoBRAIN demonstrates that our original AI can reliably identify and classify false-positive detections during colonoscopy. This high level of performance indicates robust generalizability across diverse non-polyp visual artifacts and supports the validity of this framework as an objective tool for evaluating false-positive behavior of CADe systems.
This study has several limitations. First, our sample size, while adequate to demonstrate statistical significance, remains modest, and larger multicenter validation are warranted. Second, although CAD EYE and EndoBRAIN represent leading platforms in Japan and globally, other AI systems should also be evaluated. Third, our original AI, which is currently validated using polyp-free videos, does not yet allow comprehensive performance assessment of AI systems―such as sensitivity, specificity, or false-negative rates―in mixed real-world datasets that include both polyp-positive and polyp-negative cases. Consequently, this framework is presently limited to the objective evaluation of false-positive detection. Nevertheless, we have outlined the feasibility of future validation using datasets containing polyps, which would enable extension of this framework toward more comprehensive performance assessment in real-world settings.
In conclusion, we developed the first fully automated AI system for objective FPD measurement in AI-assisted colonoscopy. This novel methodology has the potential to redefine performance standards, inform regulatory frameworks, and ultimately optimize the balance between benefit and burden in AI-assisted endoscopy.
Author contributions
Naohisa Yoshida and Xin Zhu organized the study and wrote the main manuscript text. Ruiyao Zhang and Xin Zhu finetuned original AI. Yasutaka Morimoto, Takeshi Yasuda, Reo Kobayashi, Naoto Iwai, Ken Inoue, Osamu Dohi, and Ryohei Hirose collected data. Hardesh Dhillon assisted with the language editing of the manuscript. Tomohisa Takagi supervised this study.
IRB Number
ERB-C-1704-5 (Approval date: June 5, 2024) in the Ethics Committee of Kyoto Prefectural University of Medicine
Conflict of Interest
Naohisa Yoshida, Osamu Dohi, and Tomohisa Takagi received grants for research from Fujifilm. Other authors did not have any conflict of interest.
Supplementary Files
Supplementary file 1.
Acknowledgements
We thank all members of the Department of Molecular Gastroenterology and Hepatology, Kyoto Prefectural University of Medicine for their help with this study.
Funding Statement
No funding was received for this study.
References
- 1.Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin. 2021 May; 71(3): 209-49. [DOI] [PubMed] [Google Scholar]
- 2.Corley DA, Jensen CD, Marks AR, et al. Adenoma detection rate and risk of colorectal cancer and death. N Engl J Med. 2014 Apr; 370(14): 1298-306. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Misawa M, Kudo SE, Mori Y, et al. Artificial Intelligence-Assisted Polyp Detection for Colonoscopy: Initial Experience. Gastroenterology. 2018 Jun; 154(8): 2027-9.e3. [DOI] [PubMed] [Google Scholar]
- 4.Wang P, Berzin TM, Glissen Brown JR, et al. Real-time automatic detection system increases colonoscopic polyp and adenoma detection rates: a prospective randomised controlled study. Gut. 2019 Oct; 68(10): 1813-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Repici A, Badalamenti M, Maselli R, et al. Efficacy of real-time computer-aided detection of colorectal neoplasia in a randomized trial. Gastroenterology. 2020 Aug; 159(2): 512-20. [DOI] [PubMed] [Google Scholar]
- 6.Soleymanjahi S, Huebner J, Elmansy L, et al. Artificial Intelligence-Assisted Colonoscopy for Polyp Detection : A Systematic Review and Meta-analysis. Ann Intern Med. 2024 Nov; 177(11): 1652-63. [DOI] [PubMed] [Google Scholar]
- 7.Sultan S, Murad MH, Ali MA, et al. AGA Clinical Practice Guideline on the Use of Artificial Intelligence in Colonoscopy. Gastroenterology. 2025 Feb; 168(2): 527-34. [Google Scholar]
- 8.Bretthauer M, Ahmed J, Antonelli G, et al. Use of computer-assisted detection (CADe) colonoscopy in colorectal cancer screening and surveillance: ESGE Position Statement. Endoscopy. 2025 Mar; 57(3): 667-73. [DOI] [PubMed] [Google Scholar]
- 9.Ahmad OF, Stoyanov D, Chand M, et al. Artificial intelligence in colonoscopy―false positives, false negatives, and the impact on clinical practice. Endoscopy. 2021 May; 53(5): 515-21. [Google Scholar]
- 10.Hassan C, Badalamenti M, Maselli R, et al. Computer-aided detection-assisted colonoscopy: classification and relevance of false positives. Gastrointest Endosc. 2020 Nov; 92(5): 900-4.e4. [DOI] [PubMed] [Google Scholar]
- 11.Su JR, Li Z, Shao XJ, et al. Effect of a deep-learning CADe system on adenoma detection during colonoscopy (DISCOUNT trial). Lancet Gastroenterol Hepatol. 2022 Nov; 7(11): 975-85. [Google Scholar]
- 12.Mori Y, Togashi K, Oda M, et al. Impact of artificial intelligence on colonoscopy practice: A Japanese multicenter RCT. Endoscopy. 2022 May; 54(5): 531-9. [Google Scholar]
- 13.Jocher G, Chaurasia A, and Qiu J, YOLO by Ultralytics. 2023. [Online]. Available: https://github.com/ultralytics/ultralytics
- 14.Zhang R, Yoshida N, Dohi O, et al. ARBS-YOLOv8 for Annotation Box Detection in Colonoscopy CADe Systems[C]//2024 IEEE International Conference on E-health Networking, Application & Services (HealthCom). IEEE, 2024 Nov: 1-6. [Google Scholar]
- 15.Ding X, Zhang X, Ma N, et al. RepVGG: Making VGG-style ConvNets Great Again, Proc. Of 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2021 Jun: 13728-37. [Google Scholar]
- 16.Gevorgyan Z. SIoU Loss: More Powerful Learning for Bounding Box Regression. CoRR. 2022: 2205. 1274. [Google Scholar]
- 17.East JE, Saunders BP, Burling D. AI in colonoscopy: balancing sensitivity and specificity. Nat Rev Gastroenterol Hepatol. 2020 Jul; 17(7): 429-30. [Google Scholar]
- 18.Senders JT, Arnaout O, Karhade AV. Natural and artificial intelligence in clinical decision support: a systematic review. Neurosurg Focus. 2017 Apr; 42(4): E2. [Google Scholar]
- 19.Holzwanger EA, Bilal M, Glissen Brown JR, et al. Benchmarking definitions of false-positive alerts during computer-aided polyp detection in colonoscopy. Endoscopy. 2021 Sep; 53(9): 937-40. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Hsieh YH, Tang CP, Tseng CW, et al. Computer-Aided Detection False Positives in Colonoscopy. Diagnostics (Basel). 2021 Jun; 11(6): 1113. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Misawa M, Kudo SE, Mori Y, et al. Artificial intelligence-assisted polyp detection for colonoscopy: initial experience. Gastrointest Endosc. 2018 Jan; 87(1): 166-73. [DOI] [PubMed] [Google Scholar]
- 22.Gong D, Wu L, Zhang J, et al. Detection of colorectal adenomas with a real-time computer-aided system: a multicentre randomized controlled study. Gastroenterology. 2020 Mar; 158(4): 915-25. [DOI] [PubMed] [Google Scholar]
- 23.Wang P, Berzin TM, Glissen Brown JR, et al. Real-time automatic detection system increases colonoscopic polyp and adenoma detection rates: a prospective randomised controlled study. Lancet Gastroenterol Hepatol. 2020 Apr; 5(4): 343-51. [DOI] [PubMed] [Google Scholar]
- 24.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. Gastroenterology. 2020 Aug; 159(2): 512-20. [DOI] [PubMed] [Google Scholar]
- 25.Byrne MF, Chapados N, Soudan F, et al. Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during colonoscopy using a novel computer-aided diagnostic system. Gastroenterology. 2019 May; 156(5): 1627-39. [Google Scholar]
- 26.Su JR, Li Z, Shao XJ, et al. Impact of a real-time automatic quality control system on colorectal polyp and adenoma detection: a randomized controlled trial. Gastrointest Endosc. 2020 Mar; 91(3): 406-14. [DOI] [PubMed] [Google Scholar]
- 27.Repici A, Badalamenti M, Maselli R, et al. Efficacy of real-time computer-aided detection of colorectal neoplasia in a randomized trial. Gut. 2020 Feb; 69(2): 302-10. [DOI] [PubMed] [Google Scholar]
- 28.Yamada M, Saito Y, Imaoka H, et al. Development of a real-time artificial intelligence system for colorectal polyp detection: a multicentre study. Endoscopy. 2022 Feb; 54(2): 205-12. [Google Scholar]
- 29.Spadaccini M, Hassan C, Alfarone L, et al. Comparing the number and relevance of false activations between 2 artificial intelligence computer-aided detection systems: the NOISE study. Gastrointest Endosc. 2022 Apr; 95(4): 975-81.e1. [DOI] [PubMed] [Google Scholar]
- 30.Okumura T, Imai K, Misawa M, et al. Evaluating false-positive detection in a computer-aided detection system for colonoscopy. J Gastroenterol Hepatol. 2024 May; 39(5): 927-34. [DOI] [PubMed] [Google Scholar]
- 31.Ahmad W, Yoshida N, Cardoso L, et al. Retrospective analysis of bowel bubbles during colonoscopy following bowel preparation with oral sulfate solution or 2 L polyethylene glycol-electrolyte solution with ascorbic acid. Dig Dis Sci. 2025 Jul 31. doi: 10.1007/s10620-025-09277-4. Online ahead of print. [DOI] [PubMed] [Google Scholar]
- 32.Guo Z, Nemoto D, Zhu X, et al. Polyp detection algorithm can detect small polyps: Ex vivo reading test compared with endoscopists. Dig. Endosc. 2021 Jun; 33(3): 162-9. [DOI] [PubMed] [Google Scholar]
- 33.Sugino S, Yoshida N, Guo Z, et al. Non-polypoid Colorectal Lesions Detection and False Positive Detection by Artificial Intelligence under Blue Laser Imaging and Linked Color Imaging. J Anus Rectum Colon. 2024 Jul; 8(3): 212-20. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Kobayashi R, Yoshida N, Tomita Y, et al. Detailed superiority of the CAD EYE artificial intelligence system over endoscopists for lesion detection and characterization using unique movie sets. J Anus Rectum Colon. 2024 Apr; 8(2): 61-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Yoshida N, Inoue K, Tomita Y, et al. An analysis of the function of a new artificial intelligence, CAD EYE with the lesion recognition and diagnosis for colorectal polyps in clinical practice. Int J Colorectal Dis. 2021 Oct; 36(10): 2237-45. [DOI] [PubMed] [Google Scholar]
- 36.Luo Y, Zhang J, Pagnucco M. Algorithmic vigilance: monitoring deployed AI in medicine. npj Digit Med. 2022 Jul; 15: 5: 92. [Google Scholar]
- 37.Yoshida N, Okada M, Hayashi Y, et al. Improvement of Colonoscopic Image Quality Using a New LED Endoscopic System with Specialized Noise Reduction. Diagnostics. Diagnostics (Basel). 2025 May; 15(5): 1569. [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
Supplementary file 1.