Page 33 - Read Online

George et al. Mini-invasive Surg 2024;8:4 https://dx.doi.org/10.20517/2574-1225.2023.102 Page 5 of 23

160 normal ones
[11]
Pogorelov et al. Active GI 2019 Retrospective Malaysia/ Combining colour features in RGB Training: 300 bleeding frames and SVM Sensitivity of 97.6%,
bleeding Norway and texture features for bleeding 200 non-bleeding specificity of 95.9%
detection Testing: 500 bleeding and
200 non-bleeding frames
[15] [110]
Hajabdollahi et al. Active GI 2019 Retrospective Iran Developing a low-complexity CNN Training and testing on KID CNN Sensitivity of 94.8%,
bleeding method specificity of 99.1%
Kanakatte and Active GI 2021 Prospective India Proposing compact U-Net model Training: 700 bleeding and CNN Sensitivity of 99.57%,
[16]
Ghose bleeding 700 non-bleeding specificity of 91%
Testing: 50 capsule endoscopy images
Rathnamala and Active GI 2021 Retrospective India Utilising gaussian mixture model Training: 686 bleeding and SVM Sensitivity of 99.83%,
[12]
Jenicka bleeding superpixels for bleeding detection 961 non-bleeding images specificity of 100%
Testing: 487 bleeding images and
1,160 non-bleeding images
Ghosh and Active GI 2021 Retrospective America Develop CNN-based framework for Alex-Net training: 1,410 CNN Sensitivity of 97.51%,
[17]
Chakareski bleeding bleeding identification Alex-Net testing: 940 specificity of 99.88%
SegNet training: 201
SegNet testing: 134
[18]
Ribeiro et al. Active GI 2021 Retrospective Portugal Automatic detection and Training: 820 images with red spots, CNN Sensitivity of 91.8%,
bleeding differentiation of vascular lesions 830 images with angiodysplasia/varices, specificity of 95.9%
7,620 images with normal mucosa
Testing: 206 images with red spots,
207 images with angiodysplasia/varices,
1,905 images with normal mucosa
Mascarenhas Saraiva Active GI 2022 Retrospective Portugal Create CNN-based system for Training: 10,808 images containing blood, CNN Sensitivity of 98.6%,
[19]
et al. bleeding automatic detection of blood or 6,868 with normal mucosa or other specificity of 98.9%
haematic traces in small bowel distinct pathological findings
lumen Testing; 2,702 images containing blood,
1,717 with normal mucosa or other distinct
pathological findings

Muruganantham and Active GI 2022 Retrospective India Construct dual branch CNN model Training and testing conducted on CNN No sensitivity and specificity
[20] [111]
Balakrishnan bleeding with a novel lesion attention map bleeding and Kvasir-Capsule could be found
[112]
estimator model dataset Accuracy of 94.40% for bleeding
Training: 3,430 images detection on bleeding dataset
Testing: 1,470 images Accuracy of 93.18% for ulcer,
93.89% for bleeding, 97.73% for
polyp, 96.67% for normal on
Kvasir-Capsule dataset

* [13,14]
An inconsistency was noted in publications from the same group. Jia et al.’s 2017 later work referenced the prior 2016 work but reports an inconsistent recall/sensitivity . AI: Artificial intelligence; GI:
gastrointestinal; SVM: support vector machine; CAD: computer aided design; MLP: multilayer perceptron; RGB: red, green blue; HSI: hue, saturation, intensity; BP: back propagation; PNN: probabilistic neural network;
WCE: wireless capsule endoscopy; KNN: K-nearest neighbour; CNN: convolutional neural network; ROI: region of interest; SPCH: superpixel-colour histogram; KID: koulaouzidis-iakovidis database.