Hybrid Deep Learning and Semantic Segmentation Framework for Colon Cancer Screening in Endoscopic Images

Abstract

The rising global population, lifestyle changes, and altered dietary habits have contributed to an increase in gastrointestinal diseases, including colon cancer, necessitating robust computer-aided diagnosis (CAD) systems for early detection and clinical decision-making. This paper proposes a deep semantic segmentation–assisted hybrid deep-ensemble framework (DSH-DeNet) for colon cancer screening using endoscopic images. The proposed framework applies preprocessing and UNet-based region-of-interest (ROI) segmentation to localize salient colorectal regions and enhance feature extraction from endoscopic images. Deep features are extracted using EfficientNet-B7, MobileNet-V2, and DenseNet121, followed by PCA-based feature refinement and z-score normalization. An ensemble-of-ensemble (E2E) learning strategy is employed for robust multi-class endoscopic image classification. With 98.99% accuracy, 99.24% precision, 99.17% recall, and an F-measure of 0.99, the experimental evaluation shows excellent classification performance. The results show that the suggested segmentation-assisted framework for computer-aided colon cancer screening utilizing endoscopic images is successful.