Abstract
[No abstract available]

Abstract

Abstract
Capsule endoscopy has emerged as a non-invasive alternative to traditional gastrointestinal inspection procedures, such as endoscopy and colonoscopy. Removing sedation risks, it is a patient-friendly and hospital-free procedure, which allows small bowel assessment, region not easily accessible by traditional methods. Recently, deep learning techniques have been employed to analyse capsule endoscopy images, with a focus on lesion classification and/or capsule location along the gastrointestinal tract. This research work presents a novel approach for testing the generalization capacity of deep learning techniques in the lesion location identification process using capsule endoscopy images. To achieve that, AlexNet, InceptionV3 and ResNet-152 architectures were trained exclusively in normal frames and later tested in lesion frames. Frames were sourced from KID and Kvasir-Capsule open-source datasets. Both RGB and grayscale representations were evaluated, and experiments with complete images and patches were made. Results show that the generalization capacity on lesion location of models is not so strong as their capacity for normal frame location, with colon being the most difficult organ to identify. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.