Page 47 - Read Online

Page 12 of 23 George et al. Mini-invasive Surg 2024;8:4 https://dx.doi.org/10.20517/2574-1225.2023.102

Table 4. Table of AI applications in capsule endoscopy for polyps and tumours

Year of Study
Ref. Application Study design Aim and goals Training/Validation dataset AI type Results
publication location
[55]
Li et al. Polyps and 2011 Retrospective China Utilise textural feature based Training: 450 normal samples, KNN Best Results:
tumours on multi-scale local binary 450 tumour samples MLP sensitivity of 92.33%,
pattern for tumour detection Testing: 150 normal samples, SVM specificity of 88.67%
150 tumour samples
Karargyris and Polyps and 2011 Retrospective America Utilising log Gabor filters for Polyps testing: 10 frames with polyps, SVM Ulcer detection:
[56]
Bourbakis tumours feature extraction to detect 40 normal frames sensitivity of 75.0%,
polyps and ulcers Ulcer testing: 20 ulcer frames, specificity of 73.3%
30 non-ulcer frames Polyp detection:
sensitivity of 100%,
specificity of 67.5%
Barbosa Polyps and 2012 Retrospective Portugal Extracting textural features to Dataset for training and testing: 700 tumour MLP Sensitivity of 93.9%,
[57]
et al. tumours detect polyps and tumours images, 2,300 normal images specificity of 93.1%
Mamonov Polyps and 2014 Retrospective USA/Portugal Development of binary Dataset for training and testing: BC Per frame: sensitivity
[58]
et al. tumours classifier for tumour detection 230 tumour images, 18,738 normal images of 47.4%, specificity
geometrical analysis and of 90.2%
texture content Per polyp: sensitivity
of 81.25%, specificity
of 93.47%
[59]
Liu et al. Polyps and 2016 Retrospective China Integrating multi-scale curvelet Training: WCE videos of 15 patients SVM Sensitivity of 97.8%,
tumours and fractal technology into Testing: 900 normal frames, specificity of 96.7%
textural features for polyp 900 tumour frames
detection
Yuan and Polyps and 2017 Retrospective China Construction of SSAEIM for Testing: 1,000 bubble images, 1,000 TIs, 1,000 SSAEIM Polyps: sensitivity of
[62]
Meng tumours polyp detection CIs, 1,000 polyp images 98%, specificity of
99%
Bubbles: sensitivity of
99.5%, specificity of
99.17%
TIs: sensitivity of
99%,
specificity of 100%
CIs: sensitivity of
95.5%,
specificity of 99.17%
Blanes-Vidal Polyps and 2019 Retrospective Denmark Developed algorithm to match Training: 39,550 images CNN Sensitivity of 97.1%,
[74]
et al. tumours CCE and colonoscopy polyps Testing: 8,476 images specificity of 93.3%
and construct CNN for polyp
detection
[63]
Saito et al. Polyps and 2020 Retrospective Japan Constructing CNN model for Training: 30,584 protruding lesion images CNN Sensitivity of 90.7%,
tumours protruding lesion detection Testing: 7,507 protruding lesion images, 10,000 specificity of 79.8%
normal images
[60]
Yang et al. Polyps and 2020 Retrospective China Development of algorithm Testing: 500 normal, 500 polyp images SVM Sensitivity of 95.80%,