Barrett’s esophagus (BE) is a premalignant condition associated with increased risk of esophageal adenocarcinoma (EAC) [1]. While advanced EAC carries a poor prognosis, timely detection and endoscopic treatment improves patient outcomes significantly [2]. Therefore, BE patients undergo regular endoscopic surveillance to detect neoplasia at an early stage [3]. However, due to the low progression rate to neoplasia (<1 % per patient-year) [4], its subtle mucosal changes (Fig. 1) and centralization of endoscopic treatment of neoplasia in expert centers, endoscopic detection poses a challenge for general endoscopists who conduct the bulk of routine surveillance [5,6]. Additionally, proper surveillance endoscopy requires adequate cleaning of the esophagus, extensive inspection with high definition white light endoscopy, accurate photo documentation and inspection time of over 1 min per cm of BE length [3]. The complementary Seattle protocol, dictating random 4-quadrant biopsies every 2 cm, is time consuming and is prone to sampling error [7]. Compliance to this comprehensive protocol is limited in community centers [8].

Over the last decades, several endoscopic techniques have been introduced to improve detection of early BE neoplasia. These include confocal endomicroscopy, autofluorescence imaging, volumetric laser endomicroscopy and various (virtual) chromoendoscopy techniques (e.g. acetic acid staining and narrow-band imaging) with and without magnification endoscopy. However, to date, none of these methods have shown a significant improvement in neoplasia detection rates, suffer from substantial practical challenges or are no longer commercially available [9], [10], [11], [12].

With the availability of modern high-definition endoscopes the quality of endoscopic imagery has undergone significant improvement and enables the endoscopist to fully appreciate all subtle mucosal and vascular changes. As a result, the research focus has shifted from emphasizing improved visualization to enhanced interpretation by the endoscopist. The endoscopist has thereby become the rate-limiting factor in neoplasia detection. To address this challenge and improve detection rates in the surveillance setting, endoscopists can be aided in neoplasia recognition by leveraging recent technological advancements.

Over the past ten years, artificial intelligence (AI) has gained popularity across various fields in and out of the medical domain. This can mainly be attributed to the introduction of deep learning and the remarkable increase in computational power of modern computers. Deep learning, a subfield of machine learning, employs artificial neural networks that can automatically recognize informative patterns in imagery without being hindered by human perceptual biases. Unlike endoscopists, AI is not limited by variations in intra- and inter-observer agreement or impacted by lengthy learning curves. Additionally, AI does not suffer from fatigue, which can affect endoscopists during prolonged procedures or under certain mental states. The advantages of AI’s objectivity, consistency, and tirelessness make it a promising tool to augment endoscopist performance and support real-time interpretation during endoscopic procedures.

AI algorithms have demonstrated their capability to analyze and interpret large quantities of endoscopic information, thereby supporting endoscopists in two critical aspects: neoplasia detection and lesion characterization. Computer aided detection (CADe) systems can assist endoscopists in identifying suspicious lesions that might otherwise be overlooked, potentially improving the sensitivity of neoplasia detection during endoscopy. Furthermore, computer aided diagnosis (CADx) systems can aid in lesion characterization, providing additional diagnostic insights to guide appropriate clinical decision making.

1. Download: Download high-res image (199KB)
2. Download: Download full-size image

Fig. 1. Exemplary cases of subtle, early Barrett’s neoplasia. With permission of Best Academia (best-academia.eu).