P0220Redefining treat-to-target in Crohn’s disease: a pan-intestinal, multidevice deep learning model for the detection and differentiation of ulcers and erosions in capsule endoscopy

Abstract Background Treat-to-target approaches are a mainstay in Crohn’s disease management, with evidence showcasing the benefit of capsule endoscopy driven strategies. Pan-intestinal capsule endoscopy (PCE) allows a minimally invasive evaluation of both the small bowel and colon, potentially improving disease monitoring and therapeutic decision-making. Nevertheless, this approach is hampered by lengthy reading times and observer intervariability. Artificial intelligence (AI), specifically convolutional neural networks (CNNs), have proven their ability for diagnosis of ulcers and erosions and inflammatory activity assessment. This study aimed to develop and validate an interoperable pan-intestinal AI model for detection of ulcers and erosions and determination of their clinical significance according to Saurin classification. Methods 195,955 frames (132.302 enteric, 63,653 colonic) from 1,585 exams performed in two reference centers between January 2017 and August 2025 were included for model development and testing. Data included exams from six different capsule endoscopy devices. After expert-based consensus, 71,382 frames containing ulcers and erosions were identified and differentiated according to Saurin classification, while the remaining 124 573 frames depicted normal mucosa. The dataset was divided in training / validation and testing sets with a 90%/10% patient split design. The model was evaluated using the testing using accuracy, sensitivity and specificity. Results Normal mucosa was identified with 86.6% accuracy, 88.2% sensitivity and 82.4% specificity. Ulcers and erosions were identified with 86.6% accuracy, 72.6% sensitivity and 93.0% specificity. When stratifying ulcers and erosions according to their bleeding potential, lesions with uncertain bleeding potential (P1UE) were identified with 85.6% accuracy, while ulcers with high bleeding potential (P2U) were identified with 98.6% accuracy. Conclusion PCE enables a minimally invasive treat-to-target strategy in Crohn’s disease. A multidevice AI model demonstrated high accuracy in identifying ulcers and erosions in a pan-intestinal setting, as well as stratifying them according to their clinical significance. Additionally, the pan-intestinal evaluation was performed without the need of segmentating the small bowel and colon, reducing computational requirements and representing another step toward clinical implementation. Further clinical validation studies are before integrating deep learning models into PCE, with the potential to transform patient management in inflammatory bowel disease. References: 1. Turner D, Ricciuto A, Lewis A, D’Amico F, Dhaliwal J, Griffiths AM, et al. STRIDE-II: An Update on the Selecting Therapeutic Targets in Inflammatory Bowel Disease (STRIDE) Initiative of the International Organization for the Study of IBD (IOIBD): Determining Therapeutic Goals for Treat-to-Target strategies in IBD. Gastroenterology. 2021;160(5):1570-83. 2. Ben-Horin S, Lahat A, Ungar B, Ukashi O, Yablecovitch D, Amitai MM, et al. Capsule Endoscopy-Guided Proactive Treat-to-Target Versus Continued Standard Care in Patients With Quiescent Crohn’s Disease: A Randomized Controlled Trial. Gastroenterology. 2025;169(1):85-93.e3. 3. Afonso J, Saraiva MM, Ferreira JPS, Cardoso H, Ribeiro T, Andrade P, et al. Automated detection of ulcers and erosions in capsule endoscopy images using a convolutional neural network. Med Biol Eng Comput. 2022;60(3):719-25. 4. Cardoso P, Mascarenhas M, Afonso J, Ribeiro T, Mendes F, Martins M, et al. Deep learning and minimally invasive inflammatory activity assessment: a proof-of-concept study for development and score correlation of a panendoscopy convolutional network. Therap Adv Gastroenterol. 2024;17:17562848241251569. Conflict of interest: Dr. Mendes, Francisco: Member of University of Porto spin-off Digestaid Saraiva, Miguel: Previous speaker for Medtronic, Jinshan, Jaba-Recordati Co-founder of University of Porto spin-off Digestaid Member of WGO AI taskforce Afonso, Joao: Co-founder of University of Porto spin-off Digestaid Frias, Joana: No conflict of interest Araújo, Catarina: No conflict of interest Cardoso, Pedro: Member of University of Porto spin-off Digestaid Ribeiro, Tiago: Co-founder of University of Porto spin-off Digestaid Araújo Martins, Miguel Pedro: Member of University of Porto spin-off Digestaid Almeida, Maria João: Member of University of Porto spin-off Digestaid Mota, Joana: Member of University of Porto spin-off Digestaid Pinto da Costa, António: Member of University of Porto spin-off Digestaid Andrade, Ana Patricia: Co-founder of University of Porto spin-off Digestaid Cardoso, Hélder: Co-founder of University of Porto spin-off Digestaid Mascarenhas Saraiva, Miguel: No conflict of interest Mendes, Bruno: No conflict of interest Ferreira, João: Member of University of Porto spin-off Digestaid Macedo, Guilherme: Co-founder of University of Porto spin-off Digestaid