Hybrid deep learning model for automated colorectal cancer detection using local and global feature extraction

Abstract

Colorectal cancer (CRC) ranks among the most lethal malignancies globally, underscoring the importance of timely and precise diagnosis. Although histopathological examination remains the clinical gold standard, the intricate morphology of tissue samples and inter-observer variability drive the need for robust automated methods. To address these challenges, this paper presents a hybrid deep learning model that integrates InceptionNeXt blocks, enhanced Swin Transformer blocks, and a Residual Multi-Layer Perceptron (ResMLP). In the initial stages, InceptionNeXt blocks employ multi-branch convolutions to capture nuclear morphology, glandular structures, and stromal textures, particularly benefiting limited training data scenarios. Subsequent layers utilize enhanced Swin Transformer blocks with window-based self-attention and shifted windows, effectively modeling long-range dependencies. The ResMLP component further refines feature representation via residual learning. Comprehensive evaluations on two benchmark CRC datasets—NCT-CRCHE-100K and Kather-5K—demonstrated accuracies of 99.96 % and 99.06 %, respectively, outperforming 10 state-of-the-art CNN and 10 ViT-based models. Additionally, Grad-CAM visualizations highlight the critical regions influencing classification decisions, enhancing model interpretability. These results establish the proposed method as a reliable, generalizable, and clinically viable solution for automated CRC detection.