Page 29 - Read Online

Page 2 of 23 George et al. Mini-invasive Surg 2024;8:4 https://dx.doi.org/10.20517/2574-1225.2023.102

consideration of diagnostic requirements and potential challenges is crucial. The untapped potential within vision
transformer technology hints at further evolution and even greater patient benefit.

Keywords: Artificial intelligence, capsule endoscopy, computer-assisted diagnosis, computer-assisted detection,
deep learning, vision transformer, review

INTRODUCTION
Since its inception in 2001, wireless capsule endoscopy (WCE) has revolutionised the investigation and
[1]
diagnosis of gastrointestinal (GI) diseases . However, the process of reading WCE images, along with
interpreting and diagnosing, is highly labour-intensive and error-prone considering that it is reliant on the
expertise of the reader and tens of thousands of video frames collected, of which potentially only a few
contain the lesion or pathology to be found. Hence, it is understandable that readers, with their limited
[2]
attention spans and concentration, may miss pathology or over/underdiagnose lesions which are detected .
This is why capsule endoscopy offers a “fertile” field for artificial intelligence (AI) algorithms to be
implemented, where AI can significantly streamline the reading process. Several commercial AI systems are
already available, such as Quick-View and Express-View, which can recognise potential lesions and remove
insignificant video frames. By identifying and selecting images with potential pathology for review and
removing those with no suspicion of pathology, these programs decrease the total amount of images the
reader is required to view, hence reducing overall reading time. This narrative review aimed to assess and
synthesise the current evidence on the AI applications in enhancing the capability and efficiency of capsule
endoscopy for investigation of the GI tract and propose future directions for this technology.

METHODS
Methodology for this review was formulated prior to its conduct. Ovid Embase, PubMed (incorporating
MEDLINE), and Cochrane databases were searched from database inception to 06 July 2023, with a mixture
of Medical Subject Headings (MESH) and free text terms including capsule endoscopy keywords such as
“Capsul*”, “Endoscop*”, and “Gastroscop*”, AI-related keywords such as “Artificial Intelligence”, “AI”,
“Convolutional Neural Network”, “Deep Learning”, “Computer-Assisted Diagnosis”, “Computer-Assisted
Detection”, “Transformer”, and “Vision Transformer”, and common capsule endoscopy findings such as
“Ulcer”, “Erosion”, “Vascular Lesion”, “Lesion”, “Gastrointestinal Bleed”, “Dieulafoy”, “Arteriovenous
Malformation”, “Inflammatory Bowel Disease”, “Crohn’s Disease”, “Ulcerative Colitis”, “Coeliac Disease”,
“Coeliac Sprue”, “Gluten-Sensitive Enteropathy”, “Neoplasm”, “Polyp”, “Cancer”, “Tumour”, and “Bowel
Prep”.

Study screening was conducted by three reviewers (A.G., J.K., and J.T.), with disagreements resolved
through consensus. Selection criteria were based on their relevance to the research topic of AI for capsule
endoscopy. Articles were screened for AI applications, ensuring they focused on one of the sub-categories
that were planned a priori: “Active GI Bleeding”, “Erosion and Ulcers”, “Angiodysplasia”, “Polyps and
Tumours”, “Inflammatory Bowel Disease”, “Coeliac Disease”, “Hookworm”, and “Other Applications”.
Furthermore, they were required to have constructed their own AI tool, including modalities such as
support vector machines (SVMs), Multilayer Perceptrons, and convolutional neural networks (CNNs).
Furthermore, they were screened for relevance to the field of capsule endoscopy, including domains such as
Colon Capsule Endoscopy and Small-Bowel Capsule Endoscopy. Studies were excluded if they were not in
English, were conference abstracts, did not report observational data (e.g., review articles), or did not
conform to the inclusion criteria listed above.