Page 57 - Read Online

Page 18 of 23 George et al. Mini-invasive Surg 2024;8:4 https://dx.doi.org/10.20517/2574-1225.2023.102

Table 8. Table of AI application in capsule endoscopy for bowel prep scoring
Year of Study Study Training/Validation AI
Ref. Application Aim and goals Results
publication design location dataset type
Nam Bowel prep 2021 Retrospective Korea Automatically detect Training: 500 images for CNN Sensitivity of
et al. [100] scoring and score bowel prep each score (1-5), totalling 93%,
quality on CE images 2,500 specificity of
Testing: 96 CE cases 100%
At cleansing
cut-off value of
3.25
AI: Artificial intelligence; CE: capsule endoscopy; CNN: convolutional neural network.

Table 9. Table of AI applications in capsule endoscopy for multiple lesion detection
Ref. Application Year of Study Study Aim and goals Training/Validation AI type Results
publication design
dataset
location
Park Multiple 2020 Retrospective Korea Develop CNN Training: 60,000 CNN No sensitivity or
[101]
et al. lesion model to identify significant, 60,000 specificity given;
detection multiple lesions on insignificant overall detection
CE and classify Testing: 20 CE videos rate of 81.6%
images based on
significance
[54] [110]
Xing Multiple 2020 Retrospective China Develop AGDN CAD-CAP and KID CNN Sensitivity of
[102]
etal. lesion model for WCE databases used for training 95.72% for
detection image and testing normal, 90.7% for
classification vascular images,
87.44% for
inflammatory
images
Zhu Multiple 2021 Retrospective China Construct new CAD-CAP [54] and KID [110] Deep Sensitivity of 97%
[103]
etal. lesion deep learning databases used for training neural for normal,
detection model for and testing network 94.17% for
classification and vascular images,
segmentation of 92.71% for
WCE images inflammatory
images
Guo Multiple 2021 Retrospective China Utilise CNN Training: 1,440 images CNN Sensitivity of
etal. [104] lesion models for the Testing: 360 images 96.67% for
detection automatic vascular lesions,
detection of sensitivity of
vascular and 93.33% for
inflammatory inflammatory
lesions lesions
Goel Multiple 2022 Retrospective India Develop CNN Trained and tested on CNN Sensitivity of
etal. [105] lesion framework to test collected 7,259 normal 98.06% on
detection importance of images and 1,683 collected
colour features for abnormal images database,
lesion detection Also trained and tested on sensitivity of 97%
KID [110] database on KID
Yokote Multiple 2023 Retrospective Japan Construction of Training: 17,085 images CNN Sensitivity of 91%
[106]
et al. lesion objection detection Testing: 1,396 images
detection AI model for
classification of 12
types of lesions
from CE images
Ding Multiple 2023 Retrospective China Development of AI Training: 280,426 images CNN Median sensitivity
etal. [107] lesion tool to detect Testing: 240 videos of 96.25%,
detection multiple lesion median specificity
types on CE of 83.65%
AI: Artificial intelligence; CNN: convolutional neural network; CE: capsule endoscopy; AGDN: attention guided deformation network; WCE:
wireless capsule endoscopy; CAD-CAP: computer-assisted diagnosis for capsule endoscopy; KID: koulaouzidis-iakovidis database; SVM: support
vector machine.

While AI shows high overall accuracy across many studies, it is important to note that overall accuracy