This scoping review synthesizes evidence on Artificial Intelligence (AI) for detecting small bowel (SB) lesions and neoplasia via endoscopic imaging, including capsule endoscopy and device-assisted enteroscopy. The review identified 13 studies, predominantly retrospective from China and Japan, showing high diagnostic performance (sensitivities 81.2-98.6%, specificities 88.6-99.8%) for AI systems, mostly convolutional neural networks, in automated lesion detection. While AI can improve SB lesion detection and reduce reading times, current evidence is insufficient to recommend it as a screening tool for SB cancer due to a lack of robust prospective multicenter validation.
Introduction and background
Small bowel (SB) neoplasms are rare but challenging to diagnose early due to nonspecific symptoms, complex anatomy, and lack of screening. Current diagnostic methods like capsule endoscopy and device-assisted enteroscopy face limitations such as operator dependence and prolonged interpretation times. Artificial intelligence (AI) offers a promising solution to automate image analysis, improve diagnostic performance, and reduce reading times for SB lesion detection, especially for malignant and premalignant lesions. This scoping review aims to synthesize existing literature on AI's role in detecting and characterizing these lesions in SB endoscopy.
Methods
This scoping review followed the Arksey and O’Malley, Levac, and Joanna Briggs Institute frameworks, reporting in line with PRISMA-ScR. The research question focused on AI system diagnostic performance in capsule endoscopy and device-assisted enteroscopy for detecting premalignant and malignant SB lesions. Searches were conducted in PubMed, Scopus, and Embase using terms related to AI, machine learning, deep learning, endoscopy, small bowel, and various lesion types. Studies were included if they evaluated AI on SB endoscopic images/videos and reported quantitative diagnostic metrics, excluding narrative reviews or studies without clinical image validation. Data extraction used a standardized matrix, and AI tasks were categorized as detection, classification, segmentation, monitoring, or AI-assisted reading. A descriptive narrative synthesis was performed due to study heterogeneity.
Review
The review included 13 studies, predominantly retrospective (69.2%) and originating from China and Japan. Capsule endoscopy was the most common imaging modality (69.2%). Automated lesion detection was the primary AI application (84.6%), often combined with diagnostic classification. Most models utilized convolutional neural networks (e.g., ResNet50, YOLOv5). Diagnostic performance was consistently high, with sensitivities ranging from 81.2% to 98.6%, specificities from 88.6% to 99.8%, and AUC values near 1.0. Studies highlighted AI's ability to improve sensitivity, specificity, accuracy, and significantly reduce reading times, particularly for general lesion detection rather than strictly cancer-specific applications.
Discussion
The review found that AI demonstrates high diagnostic performance in SB endoscopy, mainly for general lesion detection, with limited direct evidence for cancer-specific diagnosis. Most AI applications target various abnormalities (inflammatory, vascular, etc.), not exclusively malignant/premalignant lesions, which are fewer and morphologically diverse, making their automated characterization challenging. Capsule endoscopy is the dominant modality where AI assists detection and reduces reading time. However, the retrospective, single-center nature of most studies, reliance on selected image datasets, and image-level analyses limit generalizability and may overestimate real-world performance. While AI improves workflow efficiency, its role in cancer-specific diagnosis requires prospective multicenter validation with histopathological confirmation.
Strengths and limitations
The review followed established scoping review methodology (PRISMA-ScR) and clearly distinguished studies on malignant/premalignant lesions from other SB abnormalities, facilitating a focused interpretation. It also categorized studies by methodology, highlighting gaps. Limitations include the absence of a formal risk-of-bias assessment, high heterogeneity preventing quantitative synthesis, and the retrospective, single-center nature of most studies with image-level outcomes, which may lead to overestimated performance and limited external validity. Therefore, the findings represent an evidence map, not a confirmation of AI's clinical readiness for cancer-specific diagnosis in SB endoscopy.
Conclusions
AI shows potential in supporting small bowel abnormality detection, especially in capsule endoscopy, by reducing reading time and missed findings. However, current evidence is limited, primarily focusing on general lesion detection rather than precise neoplastic characterization or cancer-specific diagnosis. The reported high diagnostic performance may be inflated due to retrospective designs, selected image datasets, image-level analyses, and insufficient prospective multicenter validation. Consequently, the clinical implementation of AI for cancer-specific diagnosis in small bowel endoscopy requires further robust validation.
