Page 51 - Read Online

Page 14 of 23 George et al. Mini-invasive Surg 2024;8:4 https://dx.doi.org/10.20517/2574-1225.2023.102

Testing: 122 images and 97.58% on CVC-
ClinicDB databases
respectively
[77]
Lei et al. Polyps and 2023 Combined United Study is proposed to determine Study is incomplete CNN Study is incomplete
tumours prospective/retrospective Kingdom efficacy of AI tools for polyp
detection in capsule endoscopy

AI: Artificial intelligence; KNN: K nearest neighbour; MLP: multilayer perceptron; SVM: support vector machine; BC: binary classifier; WCE: wireless capsule endoscopy; SSAEIM: stacked sparse autoencoder with
image manifold constraint; TI: turbid image; CI: clear image; CCE: colon capsule endoscopy; CNN: convolutional neural network; LCDH: local colour difference; GMM: gaussian mixture model; SSMD: single shot
multibox detector; KID: koulaouzidis-iakovidis database; DBMF: dual branch multiscale feature fusion network; GI: gastrointestinal.

Current commercial endoscopes have some algorithm built to assist with interpretation. However, the training of such algorithms are based on traditional

supervised learning methods. Given the rise in higher resolution and increase the amount of training images and videos, unsupervised methods will be more
efficient and accurate.

Deep learning has shown significant promise in the field of diagnostic capsule endoscopy due to its ability to learn from large volumes of data and make
accurate predictions. Current commercial capsule endoscopes have algorithms available to assist with interpretation such as the TOP 100 feature of Rapid
Reader . However, the training of these algorithms is based on traditional supervised learning methods. Unlike traditional machine learning algorithms,
[45]
which require manual feature extraction and selection, deep learning ones can automatically learn and extract features from raw data . CNNs, in particular,
[108]
are designed to automatically and adaptively learn spatial hierarchies of features from raw data, which makes them well-suited for image classification tasks in
capsule endoscopy, as evidenced in the studies above. Given the rise in image resolution and amount of training images and video, unsupervised methods

capitalising on these AI systems will become even more efficient and accurate in future.

Despite the advantages of deep learning, it is not without its pitfalls. One of its main criticisms is the “black box” problem. Due to the complexity and depth of

these models, it can be challenging to understand and interpret how they make their predictions. This lack of transparency and interpretability can be
[109]
problematic in medical applications, where understanding the reasoning behind a diagnosis is crucial for patient care and trust . The “black box” problem
also raises concerns about the reliability and fairness of deep learning models. If the reasoning behind a model’s prediction is not clear, determining whether

the model is making decisions based on relevant features or whether it is being influenced by irrelevant or biased data can be difficult . This is an intrinsic
[109]
issue with deep learning, and hence, images must be validated prospectively prior to usage in clinical settings. Currently, AI researchers are exploring a concept
known as Explainable AI to help understand the logic and decision-making process within a black box.

When training WCE with AI, “images” obtained may not be histologically verified due to an inability to obtain biopsies without invasive enteroscopy. This
issue undoubtedly has implications for the reliability of the AI algorithms due to the potential inaccuracy of the training dataset used. This may adversely affect

the diagnostic accuracy, causing either false-positive or false-negative results, both of which have significant clinical implications. The issue of data quality can