10.1055/a-2695-2841
We appreciate the Letter to the Editor by Deding et al. [1]: “Urgency for standardized protocols to improve clinical implementation of artificial
intelligence (AI) in endoscopic diagnostics”, emphasizing the need for development
of AI to follow protocols rather than being fragmented, as summarized in the systematic
review [2] the Letter to the Editor addressed [1]. Their emphasis on AI in capsule endoscopy (CE) is timely, especially given the
European Union’s support of initiatives such as I-Supported Image Analysis in Large Bowel
Camera Capsule Endoscopy (AICE) and in general toward improvements in diagnosing and
treating colorectal cancer. CE interpretation remains a labor-intensive task with
high interobserver variability. An Endoscopy International Open study suggested that
learning small bowel CE may be more difficult and labor-intensive than previously
assumed, because none of 22 gastroenterologists reached a learning plateau with sufficient
competencies after reviewing 20 small bowel CE, with an accumulated specificity for
diagnosis of just 63% and sensitivity of just 65% [3].
We acknowledge and agree with their concerns regarding the limitations of human-centric
reference standards such as the Boston Bowel Preparation Scale (BBPS) to train AI,
which is widely used yet inconsistently correlated with clinically relevant outcomes
such as adenoma detection rate (ADR), polyp detection rate (PDR), and adenoma miss
rate (AMR) [2]. Importantly, one of the eight studies included in our review used a fecal-to-mucosa
pixel ratio and was validated against > 1,400 colonoscopies and an external dataset,
correlating with PDR rather than just BBPS [4]. In addition, it was the only AI to be open-source, allowing for external validation,
as an important part of protocol for validating AI, highlighted by Deding et al. [1].
In parallel with AICE, through the intelligent robotic endoscopy (IRE) initiative
(https://ire4health.eu/), we have published a freely available dataset of over 1,400 clinical colonoscopies
and 100 simulated colonoscopies with full colonoscope positional tracking throughout
the procedure [5]. This dataset facilitates development of AI systems that incorporate spatial-temporal
tracking, particularly relevant for development of new modalities such as robotic
endoscopy, through IRE.
In conclusion, we concur with Deding et al. [1] that future AI models should be explainable and validated against hard clinical
outcome measures such as ADR, PDR, or AMR, aligning with the recent European Society
of Gastrointestinal Endoscopy position statement on the expected value of AI in endoscopy
[6]. To that end, reporting guidelines such as Quality assessment of AI preclinical
studies in diagnostic endoscopy (QUAIDE) [7] represent a major step forward as a protocol for standardization. We support widespread
adoption of such frameworks to ensure standardization, reproducibility, and meaningful
clinical implementation of AI in both conventional colonoscopy and CE along with making
these AI algorithms and datasets open-source for external validation and training.
Bibliographical Record
Kristoffer Mazanti Cold, Amaan Ali, Lars Konge, Flemming Bjerrum, Laurence Lovat,
Omer Ahmad. Author reply to letter to the editor: From fragmentation to frameworks:
Standardizing AI in gastrointestinal endoscopy. Endosc Int Open 2025; 13: a26952884.
DOI: 10.1055/a-2695-2884
