Page 37 - Read Online

George et al. Mini-invasive Surg 2024;8:4 https://dx.doi.org/10.20517/2574-1225.2023.102 Page 7 of 23

Table 2. Table of AI applications in capsule endoscopy for erosions and ulcers

Year of Study
Ref. Application Study design Aim and goals Training/Validation dataset AI type Results
publication location
[24]
Li and Meng Erosions and 2009 Retrospective China Utilising chromaticity moment to Training: 1,350 normal samples and MLP Bleeding: sensitivity of 87.81%,
ulcers discriminate normal regions and 1,350 abnormal samples specificity of 88.62%
abnormal region Testing: 450 normal samples and Ulcer: sensitivity of 84.68%,
450 abnormal samples specificity of 92.97%
[23]
Charisis et al. Erosions and 2010 Retrospective Greece Using BEEMD to extract intrinsic Dataset: 40 normal and 40 ulcerous SVM Sensitivity of 95%, specificity
ulcers mode functions images of 96.5%
90% for training, 10% for testing
[25]
Charisis et al. Erosions and 2012 Retrospective Greece Associate colour with structure 87 normal images, 50 “easy ulcer case” MLP and SVM: sensitivity of 98.9%,
ulcers information in order to discriminate images, 37 “hard ulcer case” images SVM specificity of 96.9%, for “easy
between healthy and ulcerous 90% was used for training, 10% for ulcer”; sensitivity of 95.2%,
tissue testing specificity of 88.9%, for “hard
ulcer”
MLP: sensitivity of 94.6%,
specificity of 98.2%, for “easy
ulcer”; sensitivity of 82%,
specificity of 95.1%, for “hard
ulcer”
Iakovidis and Erosions and 2014 Retrospective Greece/ Derive colour feature-based pattern Training: 1,233 images SVM Sensitivity of 95.4%, specificity
[26]
Koulaouzidis ulcers United recognition method Testing: 137 images of 82.9%
Kingdom
[27]
Fan et al. Erosions and 2018 Retrospective China Automatic erosion detection via Ulcer training: 2,000 ulcer images, CNN Ulcers: sensitivity of 96.8%,
ulcers deep neural network 2,400 normal images specificity of 94.79%
Ulcer testing: 500 ulcer images, Erosions: sensitivity of 93.67%,
600 normal images specificity of 95.98%
Erosion training: 2,720 ulcer images,
3,200 normal images
Erosion testing: 690 ulcer images,
800 normal images
[28]
Khan et al. Erosions and 2019 Retrospective Pakistan Utilising DenseNet CNN for Training: 2,800 ulcers, 2,800 bleeding, MLP Sensitivity of 99.40%,
ulcers stomach abnormality classification and 2,800 healthy regions specificity of 99.20%
Testing: 1,200 ulcers, 1,200 bleeding,
and 1,200 healthy regions
[30]
Wang et al. Erosions and 2019 Retrospective China Use deep convolutional neural Training: 15,781 ulcer frames and CNN Sensitivity of 89.71%,
ulcers networks to provide classification 17,138 normal frames specificity of 90.48%
confidence score and bounding box Testing: 4,917 ulcer frames and
marking area of suspected lesion 5,007 normal frames
[31]
Aoki et al. Erosions and 2019 Retrospective Japan Develop CNN system based on a Training: 5,360 ulcer and erosion images CNN Sensitivity of 88.2%, specificity
ulcers single shot multibox detector Testing: 440 ulcer and erosion images, of 90.9%
10,000 normal images
[32]
Ding et al. Erosions and 2019 Retrospective China Characterise SB-CE images as Training: 158,235 images from CNN Sensitivity of 99.90%,
ulcers multiple lesion types 1,970 patients specificity of 100%
Testing: 113, 268, 334 images from