Page 42 - Read Online

Page 10 of 23 George et al. Mini-invasive Surg 2024;8:4 https://dx.doi.org/10.20517/2574-1225.2023.102

Table 3. Table of AI applications in capsule endoscopy for vascular lesions and angiodysplasias

Year of Study Study AI
Ref. Application Aim and goals Training/Validation dataset Results
publication design location type
Gan Vascular lesions and 2008 Retrospective China Develop computer-aided screening and Dataset of 236 patients with lesion, IPS Median sensitivity of 74.2%
[44]
et al. angiodysplasias diagnosis for enteric lesions in CE and 86 without lesion for training and
validation
Leenhardt Vascular lesions and 2018 Retrospective France Utilise CNN for detection of AGD in SB-CE Training: 300 normal frames, CNN Sensitivity of 100%, specificity
[54]
et al. angiodysplasias images 300 AGD frames of 96%
Testing: 300 normal frames,
300 AGD frames

Arieira Vascular lesions and 2019 Retrospective Portugal Evaluate accuracy and efficacy of “TOP 100” Testing: 97 patients TOP No sensitivity or specificity.
[45]
et al. angiodysplasias feature 100 Accuracy of 83.5% for P2
lesions,
95.5% for AGD, 56.7% for
ulcers,
100% for active bleeding sites
Vieira Vascular lesions and 2019 Retrospective Portugal Automatic detection of AGD in WCE videos Dataset: 27 images from KID MLP MLP: sensitivity of 96.60%,
[46] [110]
et al. angiodysplasias database , and specificity of 94.08%
additional 248 AGD images, SVM SVM: sensitivity of 96.58%,
550 normal images specificity of 92.24%
Vezakis Vascular lesions and 2019 Retrospective Greece Combining of low-level image analysis, feature Training: 350 normal images, CNN Sensitivity of 92.7%, specificity
[47]
et al. angiodysplasias detection, and machine learning for AGD 196 bubble images, of 99.5%
detection in WCE images 75 blood vessel images,
104 AGD images
Testing: 3 full-length WCE
Leenhardt Vascular lesions and 2019 Retrospective France Develop CNN methodology to detect GIA in Training: 300 normal frames, CNN Sensitivity of 100%, specificity
[48]
et al. angiodysplasias SB-CE 300 GIA frames GIA of 96%
Testing: 300 normal frames,
300 GIA frames
Tsuboi Vascular lesions and 2020 Retrospective Japan Development of CNN system based on SSMB Training: 2,237 angiodysplasia images CNN Sensitivity of 98.8%, specificity
[49]
et al. angiodysplasias for small bowel AGD detection Testing: 488 AGD images, of 98.4%
10,000 normal images
Aoki Vascular lesions and 2021 Retrospective Japan Construct CNN based system for various Training: 44,684 images of CNN No sensitivity or specificity
[50]
et al. angiodysplasias abnormality detection abnormalities and 21,344 normal reported. Accuracy of 100% for
images mucosal breaks,
Testing: 379 full small-bowel CE 97% for AGD, 99% for
videos protruding lesions,
and 100% for blood content
Hwang Vascular lesions and 2021 Retrospective Korea Develop CNN algorithm for categorisation of Training: 11,776 haemorrhagic lesions, CNN Sensitivity of 97.61%,
[51]
et al. angiodysplasias SBCE videos into haemorrhagic lesions and 18,448 ulcerative lesions, specificity of 96.04%
ulcerative lesions 30,224 normal images
Testing: 5,760 images
Hosoe Vascular lesions and 2022 Retrospective Japan Detect common findings on SBCE images using Training: 33 SBCE cases CNN Sensitivity of 93.4%, specificity
[52]
et al. angiodysplasias CNN framework with aim to reduce false- Testing: 35 SBCE cases of 97.8%