Page 32 - Read Online

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

Table 1. Table of AI applications in capsule endoscopy for active GI bleeding

Year of Study Study
Ref. Application Aim Training/Validation dataset AI type Results
publication design location
[3]
Giritharan et al. Active GI 2008 Retrospective America Develop a method to re-balance 550 bleeding images SVM Sensitivity of 80%
bleeding training images
[8]
Li and Meng Active GI 2009 Retrospective China Develop new CAD system utilising Training: 1,800 bleeding patches and MLP Sensitivity of 92.6%,
bleeding colour-texture features and neural 1,800 normal patches specificity of 91%
network classifier Testing: 1,800 bleeding patches and
1,800 normal patches
[4]
Pan et al. Active GI 2009 Retrospective China Use colour-texture features in RGB Training: 10,000 pixels BP neural Sensitivity of 93%,
bleeding and HSI as input in BP neural Testing: 3,172 bleeding images and network specificity of 96%
network 11,458 non-bleeding images
[7]
Pan et al. Active GI 2011 Retrospective China Use colour-texture features in RGB Training: 50,000 pairs PNN Sensitivity of 93.1%,
bleeding and HSI as input in PNN Testing: 3,172 bleeding images and specificity of 85.8%
11,458 non-bleeding images
[5]
Ghosh et al. Active GI 2014 Retrospective Bangladesh Use RGB colour-texture feature in Training: 50 bleeding images and SVM Sensitivity of 93.00%,
bleeding SVM 200 non-bleeding images specificity of 94.88%
Testing: 400 bleeding and
1,600 non-bleeding images
[6]
Hassan and Haque Active GI 2015 Retrospective Bangladesh Utilise characteristic patterns in Training: 600 bleeding and SVM Sensitivity of 99.41%,
bleeding frequency spectrum of WCE 600 non-bleeding frames specificity of 98.95%
images Testing: 860 bleeding and
860 non-bleeding images
[9]
Yuan et al. Active GI 2016 Retrospective China Construct two-fold system for Testing: 400 bleeding frames and SVM and Sensitivity of 92%,
bleeding detection and localisation of 2,000 normal frames KNN specificity of 96.5%
bleeding regions
[13] *
Jia and Meng Active GI 2016 Retrospective China Develop deep neural network that Training: 2,050 bleeding and CNN Sensitivity of 99.20%
bleeding can automatically and 6,150 non-bleeding images
hierarchically learn high-level Testing: 800 bleeding,
features 1,000 non-bleeding
[14]
Jia and Meng Active GI 2017 Retrospective China Combine handcrafted and CNN Training: 200 bleeding frames and CNN Sensitivity of 91%
bleeding features for characterisation 800 normal frames
Testing: 100 bleeding frames and
400 normal frames
[21]
Kundu et al. Active GI 2018 Retrospective Bangladesh Detecting bleeding images based Testing: 5 videos, with 100 image KNN Sensitivity of 85.7%,
bleeding on precise ROI detection in frames each specificity of 69.6%
normalised RGB colour plane
[10]
Ghosh et al. Active GI 2018 Retrospective Bangladesh/ Utilising cluster-based statistical Testing: 5 WCE videos SVM Sensitivity of 96.5%,
bleeding Canada feature extraction for global feature specificity of 94.6%
vector construction
[22]
Xing et al. Active GI 2018 Retrospective China Using SPCH feature based on the Training: 340 bleeding frames and KNN Sensitivity of 98.5%,
bleeding principal colour spectrum to 340 normal ones specificity of 99.5%
discriminate bleeding frames Testing: 160 bleeding frames and