Page 17 - Read Online

Tsuboi et al. Mini-invasive Surg 2024;8:26 https://dx.doi.org/10.20517/2574-1225.2023.94 Page 11 of 21

passive procedure without the added stimuli of instrument manipulation or patient interaction. In addition,
because CE automatically captures images that move physiologically, the entire lesion may be difficult to
visualize or may only be visible at the edges of the image. These factors may increase the risk of missed
lesions during the physicians’ reading process. To address this concern, assistive reading technologies have
been incorporated into the reading software for each CE model.

TM
For instance, QuickView mode has been included in RAPID software since ver. 6.0. It is a function that
automatically extracts images with a high probability of abnormal findings using a specific algorithm.
Although the use of QuickView mode is expected to reduce the number of images read and shorten the
reading time, it is not recommended for primary reading, as many reports indicate that it is not sensitive
TM
enough to detect abnormal findings [119-122] . In addition to Quick view mode, RAPID software ver. 9.0
introduced the “TOP 100” feature, which automatically selects the 100 images most likely to contain
abnormalities. Studies have demonstrated the effectiveness of such assistive features. Arieira et al. reported
that TOP 100 correctly identified all sites of active bleeding and detected a majority of significant lesions
(83.5%), including angioectasia, a frequent source of small-bowel bleeding, with 95% accuracy .
[123]
Gomes et al. evaluated the Express-View mode for MiroCam®, which achieved a diagnostic accuracy of
91% . The per-patient sensitivity was 83.1% for all clinically significant lesions and 56.2% for all lesions.
[124]
For the OMOM capsules, the images were reviewed using Vue Smart Software (Jinshan Science &
Technology Co.). The Vue Smart Software system uses AI-based diagnostic assistance technologies. It
deletes up to 90% of the captured images and automatically selects images of suspected abnormal lesions.

Computer-aided diagnosis (CAD) combines AI, computer vision, and pathology image processing to
automatically detect abnormalities and assist physicians in providing more accurate diagnoses.
Convolutional neural networks, a form of deep learning, are highly beneficial in endoscopy [125-128] .

Numerous studies have reported the usefulness of AI in CE. Various studies have reported the automatic
[129]
detection of lesions, such as erosions, ulcers [The area under the receiver operating characteristic curve
Receiver Operating Characteristic - Area Under the Curve (ROC-AUC): 0.958, sensitivity: 88.2%,
specificity: 90.9%], angioectasia (ROC-AUC: 0.998, sensitivity: 98.8%, specificity: 98.4%), blood
[130]
[131]
(ROC-AUC: 0.9998, sensitivity: 96.6%, specificity: 99.9%), and various protruded lesions (ROC-AUC:
[132]
0.911, sensitivity: 90.7%, specificity: 79.8%). AI assistance has also been shown to impact reading
[133]
[134]
time and has been compared with assistive systems, including RAPID software . Aoki et al.
reported that the reading time of SBCE after the first screening of AI reading was significantly shorter
[133]
compared to only physicians reading (expert: 3.1 min vs. 12.2 min, trainee: 5.2 min vs. 20.7 min) . Aoki
et al. evaluated the detection capability of their construction AI system compared to Quick View mode
[134]
equipped with RAPID software . They reported that the detection rate of the AI system was significantly
higher than QuickView mode (99% vs. 89%, respectively). The reading time of AI systems is astonishingly
fast. Reports indicate that the use of AI for reading can significantly reduce the reading time [134-140]. This
reduction in reading time is particularly pronounced for trainees, and it also alleviates the
psychological stress associated with the reading process . SBCE reading with AI assistance could maintain
[141]
a diagnostic yield comparable to that of expert physicians and reduce the reading time. In a small-bowel
follicular lymphoma assessment, AI has proven useful for disease progression evaluation .
[142]

Furthermore, the application of AI in CCE has been reported. AI has shown promise in detecting colorectal
polyps [143,144] (ROC-AUC: 0.902-0.97, sensitivity: 79.0-90.7%, specificity: 87.0-92.6%), erosions, ulcers [145]
(ROC-AUC: 1.00, sensitivity: 96.9%, specificity: 99.9%), and blood (sensitivity: 97.2%, specificity: 99.9%)
[146]
in the colon. In CCE, the reading time can be lengthy, and AI has been demonstrated to reduce this reading
[147]
time while maintaining high sensitivity .