ConCapsNet: Efficient Detection and Classification of Colorectal Cancer Using Convolutional Capsule Network

In the world, CRC is the leading cause of cancerrelated mortality, so diagnostic systems need to be accurate and efficient. This study presents ConCapsNet, a hybrid Convolutional Capsule Network (CapsNet) framework, designed to enhance the detection and classification of colorectal cancer using the Kvasir dataset. ConCapsNet overcomes the limitations of traditional convolutional neural networks (CNNs), which often struggle to capture global, hierarchical relationships necessary for detecting irregularly shaped polyps and lesions in gastrointestinal endoscopy images. In this framework, lesion-aware feature extraction is provided using CNNs, and classification is provided through capsule networks. The CNN module extracts mid-level features while preserving fine-grained lesion details, which are subsequently passed to the Capsule Network. This network models spatial relationships and ensures robustness against image deformations, such as stretching and rotation. Trained and validated using the Kvasir dataset, designed for automatic endoscopy image analysis, ConCapsNet achieved remarkable performance, with an overall accuracy of 99.3%, sensitivity of 98.9%, and specificity of 99.1%, while reducing inference time by 15% compared to traditional CNN architectures. It is effective in detecting both early-stage and advanced CRC, with minimal false positives, making it a suitable tool for real- time clinical diagnosis. ConCapsNet demonstrates its potential as a robust, efficient, and scalable solution for automated colorectal cancer diagnosis.