Page 30 - Read Online

George et al. Mini-invasive Surg 2024;8:4 https://dx.doi.org/10.20517/2574-1225.2023.102 Page 3 of 23

Figure 1. Modified PRISMA flow diagram of search strategy and study selection process.

SEARCH RESULTS
In our search, 824 articles were retrieved, of which 291 duplicates and 31 abstracts were removed. After
study screening and full-text review, 106 articles were included for analysis in the present review. Data was
synthesised into tabular and narrative formats. For studies with multiple trials, the best result achieved by
the models was used.

Additionally, we have designed a modified PRISMA flow-chart [Figure 1].

RESULTS
Active GI bleeding
Automatic haemorrhage detection is one of the largest researched applications of AI for capsule endoscopy.
[3-7]
From machine learning models such as the SVM and probabilistic neural network (PNN) methods , the
field has progressed into deep learning models with enhanced efficacy and accuracy. Other models utilising
[8]
[4]
multi-layer perceptrons (MLP) and back-propagation neural networks have also been replaced with deep
learning, with this shift appearing to primarily have occurred post-2016. Only four SVM-based models [9-12]
were constructed following 2016, compared with eight CNN models [13-20] and two Kernel Neural
Networks [21,22] . For example, in 2021, Ghosh et al. constructed a CNN-based deep learning framework via
the CNN architecture AlexNet, achieving a sensitivity of 97.51% and specificity of 99.88%, significantly
enhanced from the sensitivity of approximately 80% previously mentioned by Girithiran et al. [3,17] . However,
SVM models such as that of Rathnamala et al. in 2021 also produced excellent results, with a sensitivity of
[12]
99.83% and specificity of 100% reported . More recently, in 2022, Mascarenhas Saraiva et al. constructed a
CNN detecting blood and haematic residues in the small blood lumen with a sensitivity and specificity of
[19]
98.6% and 98.9%, respectively, with an impressive speed of around 184 frames/s . Based on the current
literature for gastrointestinal haemorrhage, incorporating AI significantly improves investigative capability.
However, further implementation work is necessary to optimise its accuracy [Table 1].