The recent systematic review by Cold et al. [1] offers a comprehensive review of artificial intelligence (AI) tools designed for
preparation scoring in colonoscopy. Despite confirming high performance, the review
highlights significant limitations such as inconsistent validation methods, a lack
of external reproducibility, minimal integration into clinical workflows, and suboptimal
reference standards. Most importantly, correlation with clinically meaningful endpoints
such as adenoma detection rate (ADR) or adenoma miss rate (AMR) is scarce, raising
concerns about practical impact on patient outcomes. This is particularly problematic
when the benchmark itself (colonoscopy) has a mean AMR of 26%, raising serious concerns
about its suitability as a reference standard for validating AI systems, especially
when those systems are applied to other modalities such as capsule endoscopy (CE),
where the diagnostic context is fundamentally different.
Although interobserver agreement using the Boston Bowel Preparation Score (BBPS) may
be high, it holds little clinical value if not linked to hard outcomes. Moreover,
training AI on only high-consensus images introduces bias, making models less effective
in ambiguous or difficult cases. Although it streamlines annotation, it biases AI
toward clear-cut cases, reducing performance in diagnostically challenging images.
Clinical-grade AI must be trained on both ambiguous and consensus cases to ensure
balanced diagnostic capability.
Nadimi et al. [2] address these gaps with an explainable AI (XAI) model integrated into the colon
CE (CCE) workflow, automating polyp detection, characterization, and sizing. Using
methods such as GradCAM++, CartoonX, and Pixel RDE, the system combines technical
strength with transparency to support clinical trust and adoption. Their work highlights
the value of combining AI tools to achieve more precise and detailed results, aligning
with the discussion by Cold et al. In parallel, Mascarenhas-Saraiva et al. [3] developed a rule-based AI for bowel prep assessment in CCE, prioritizing interpretability
and real-time feedback. Designed for clinical use, it aims to enhance trust and reduce
fatigue or variation in high-volume endoscopy settings.
Most current AI models in capsule endoscopy ignore key clinical cues such as frame
and pixel clustering and time spent on specific regions. This lack of temporal modeling
limits diagnostic realism and disconnects AI behavior from clinician reasoning. Although
less critical in live colonoscopy, it is essential in asynchronous CCE. In addition,
Moen et al. [4] outlined critical challenges and future directions for AI in CCE. Their review emphasizes
the need for harmonized datasets, transparent AI models, and coordinated clinical
validation. These systemic challenges support the rationale for a unified framework
such as The Artificial Intelligence in Capsule Endoscopy (AICE) project to ensure
alignment between technological development and clinical implementation.
AICE was funded by EU to address precisely these issues. AICE is a European collaboration
committed to building a cohesive and clinically driven framework for AI implementation
in CCE. It aims to standardize evaluation protocols, support multicenter validation
efforts, and advocate for alignment with established international standards, including
those of the European Society of Gastrointestinal Endoscopy and the World Endoscopy
Organization. By fostering collaboration between clinicians, data scientists, and
regulatory stakeholders, AICE seeks to transform AI from a set of isolated tools into
an integrated component of endoscopic care. AICE also prioritizes explainability and
usability as core pillars of responsible AI. The initiative supports use of interpretable
models, transparent reporting guidelines, and validation in real-world settings. This
approach ensures that end-users, especially clinicians, can trust and effectively
interact with AI outputs. It also lays the foundation for regulatory approval and
integration into national screening programs, particularly as health systems become
more digitally mature.
However, despite notable individual advances, the broader field of AI in endoscopy
remains fragmented. There is still limited consensus on validation protocols, outcome
measures, reference standards, and implementation standards [5]. It looks as if the momentum of innovation has outpaced the frameworks for adoption.
What is needed now is alignment on reporting, external validation across populations,
and endpoints that reflect real patient outcomes, including diagnostic uncertainty,
not just retrospective accuracy.
Bibliographical Record
Ulrik Deding, Benedicte Schelde-Olesen, Ervin Toth, Anastasios Koulaouzidis. Urgency
for standardized protocols to improve clinical implementation of artificial intelligence
in endoscopic diagnostics. Endosc Int Open 2025; 13: a26952841.
DOI: 10.1055/a-2695-2841
