Page 54 - Read Online

Page 16 of 23 George et al. Mini-invasive Surg 2024;8:4 https://dx.doi.org/10.20517/2574-1225.2023.102

625 normal mucosa images
[88]
Majtner et al. Inflammatory 2021 Retrospective Denmark Detection and classifying CD lesions based on Training: 5,419 images CNN Sensitivity of 96.2%,
bowel disease severity Testing: 1,558 images specificity of 100%
[89]
Ferreira et al. Inflammatory 2022 Retrospective Portugal Automatically detecting ulcers and erosions in Training: 19,740 images CNN Sensitivity of 90%,
bowel disease the small intestine and colon Testing: 4,935 images specificity of 96%
[90]
Higuchi et al. Inflammatory 2022 Retrospective Japan Classifying ulcerative colitis lesions using MES Training: 483,644 images CNN No Sensitivity or
bowel disease criteria Testing: 255,377 images specificity given.
Accuracy of 98.3% on
validation
[91]
Kratter et al. Inflammatory 2022 Retrospective Israel Accurately identify ulcers on capsule endoscopy Database of 15,684 normal mucosa CNN No Sensitivity or
bowel disease by combining algorithm viable for two models of images, specificity given.
capsule endoscope 17,416 ulcerated mucosa images used Accuracy of 97.4% on
for training and validation validation
Mascarenhas Inflammatory 2023 Retrospective Portugal Construct CNN for automatic classification of Database of 6,844 normal mucosa CNN Sensitivity of 97.4%,
[92]
et al. bowel disease various types of pleomorphic gastric lesions images, specificity of 95.9%
1,407 protruding lesion images,
994 ulcer and erosion images,
822 vascular lesion images,
2,851 haematic residue images used for
training and validation

AI: Artificial intelligence; WCE: wireless capsule endoscopy; CED: canny edge detector; SVM: support vector machine; SB: small bowel; CNN: convolutional neural network; RANN: recurrent attention neural network;
CE: capsule endoscopy; NSAID: non-steroidal anti-inflammatory drugs; MES: mayo endoscopic subscore.

be mitigated by ensuring that the AI models are trained on high-quality, histologically proven images, such as the French-created CAD-CAP. This could
involve collaborations with medical institutions and experts to curate and verify the training datasets.

The current AI models used in capsule endoscopy also do not appear to harness the potential of vision transformers (ViTs), a state-of-the-art AI model
adapted from natural language processing, which utilises self-attention methods for training. ViTs offer a far superior capacity for data handling compared to

other deep learning models, with approximately four times as much capacity as that of traditional CNNs. Moreover, their ability to combine spatial analysis
with temporal analysis allows them to demonstrate a much superior performance in image-based tasks. Their employment in capsule endoscopy could open
the door to more precise lesion characterisation, thereby enhancing the diagnostic potential of this technology. The lack of current models using ViTs presents
a notable gap in the field. However, this is primarily due to the recency of the technology in the medical imaging world. The use of ViT in endoscopy has only

been explored very recently in research settings, and more applications are expected in the near future.

However, the potential of AI-assisted capsule endoscopy, particularly for colonoscopy for polyp detection and characterisation, is notable. While capsule

endoscopy is quite costly compared to the Faecal Occult Blood Test (FOBT), it could serve as an alternative for patients where FOBT may yield high false