ROLE OF ARTIFICIAL INTELLIGENCE IN SMALL BOWEL CAPSULE ENDOSCOPY TRAINING

Aims Reading of Small Bowel (SB) capsule videos still represents the main limitation since it requires time and prolonged attention, even more for novices. Artificial Intelligence (AI) in small bowel CE might represent a key strategy either in routine clinical use or in a training setting. Primary aim of this study was to measure the inter-observer agreement among novices and experts performing standard or AI-assisted reading. Secondary aim was to evaluate reading time in both modalities.

Methods 10 videos of patients who performed SB CE (Navicam, Ankon, China) for suspected SB bleeding from July to September 2021 were retrospectively evaluated by 2 experts (>500 cases) and 4 novices (<5 cases). One expert and 2 novices were radomized to blindly review videos in standard modality (SR) or with the assistance of AI. Findings were classified according to the Saurin classification. The agreement between experts and novices was evaluated for P2 lesions in a per-patient analysis using Cohen's kappa statistic.

Results Of 10 SB CE videos evaluated at per-patient analysis, expert readers reported the same main diagnoses (100% inter-observer agreement) whereas novices showed moderate to substantial agreement when compared to experts. Mean reading time using AI resulted significantly lower (p<0.005, 95% IC) for both experts and novices (see the Table below).

Table 1

	Cohen's k	Mean reading time+/- SD
Expert – SR	1	43.7+/- 10.97
Expert – AI	1	4+/- 1.49
Novice 1 – SR	1	42.1+/- 9.8
Novice 2 – SR	0.58	40+/- 6.6
Novice 1 – AI	0.61	14+/- 4.35
Novice 2 – AI	0.61	10.1+/- 2.96

Conclusions In a training setting, these preliminary data suggest that artificial intelligence significantly reduces the reading time of non-expert readers without affecting the overall accuracy and the inter-observer agreement.

Publication History

Article published online:
14 April 2022