Frame Classification in Video Capsule Endoscopy Using an Improved Capsule Network

The non-invasive technique of Video Capsule Endoscopy (VCE) enables a thorough examination of the human intestinal tract, generating numerous images from various segments of the gastrointestinal system. Since manual image analysis is labor-intensive and time-consuming, there is a high demand for automated diagnostic systems to identify digestive disorders from VCE frames. Most existing studies rely on Convolutional Neural Networks (CNNs). These approaches demonstrate limited efficacy in capturing spatial relationships between features and recognizing objects across various poses and variations within VCE images. Therefore, this paper proposes an improved capsule network to increase the preservation and comprehension of spatial connections between features, hence improving the recognition of patterns and objects that have been rotated, scaled, or otherwise transformed. The proposed network employs convolutional-capsule layers with a modified routing algorithm between capsules to encode and extract characteristics, resulting in part-whole relationships. Furthermore, the network applies vector length calculations to feature vectors obtained via the layers mentioned above to determine the classification of images as normal or abnormal. The proposed network attains state-of-the-art performance, with a notable accuracy of 96.78% on the public Kvasir-Capsule dataset, while requiring considerably fewer parameters than competing classification networks.