AI-assisted capsule endoscopy reading in suspected small bowel bleeding: a multicentre prospective study

作者：Cristiano Spada, Stefania Piccirelli, Cesare Hassan, Clarissa Ferrari, Ervin Tóth, Begoña González-Suárez, Martin Keuchel, M E McAlindon, Ádám Finta, A Rosztóczy, Xavier Dray, Daniele Salvi, Maria Elena Riccioni, Robert Benamouzig, Amit Chattree, Adam Humphries, Jean‐Christophe Saurin, Edward J. Despott, Alberto Murino, Gabriele Wurm Johansson, Antonio Giordano, Peter Baltes, Reena Sidhu, M Szalai, Krisztina Helle, Artúr Németh, T Nowak, Rong Lin, Guido Costamagna · 发表于：The Lancet Digital Health · 年份：2024 · DOI：10.1016/s2589-7500(24)00048-7 · 被引用次数：56 · 研究领域：Gastrointestinal Bleeding Diagnosis and Treatment、Colorectal Cancer Screening and Detection、Gastrointestinal motility and disorders

BACKGROUND: Capsule endoscopy reading is time consuming, and readers are required to maintain attention so as not to miss significant findings. Deep convolutional neural networks can recognise relevant findings, possibly exceeding human performances and reducing the reading time of capsule endoscopy. Our primary aim was to assess the non-inferiority of artificial intelligence (AI)-assisted reading versus standard reading for potentially small bowel bleeding lesions (high P2, moderate P1; Saurin classification) at per-patient analysis. The mean reading time in both reading modalities was evaluated among the secondary endpoints. METHODS: Patients aged 18 years or older with suspected small bowel bleeding (with anaemia with or without melena or haematochezia, and negative bidirectional endoscopy) were prospectively enrolled at 14 European centres. Patients underwent small bowel capsule endoscopy with the Navicam SB system (Ankon, China), which is provided with a deep neural network-based AI system (ProScan) for automatic detection of lesions. Initial reading was performed in standard reading mode. Second blinded reading was performed with AI assistance (the AI operated a first-automated reading, and only AI-selected images were assessed by human readers). The primary endpoint was to assess the non-inferiority of AI-assisted reading versus standard reading in the detection (diagnostic yield) of potentially small bowel bleeding P1 and P2 lesions in a per-patient analysis. This stu...