The rising incidence of colorectal cancer necessitates accurate diagnostic tools to support pathologists. This study proposes a hybrid deep learning framework combining DenseNet-121 and EfficientNet-B3 for enhanced feature extraction and classification, achieving 99.63% accuracy. Unlike traditional models, it utilizes adaptive feature fusion and multi-head attention to boost feature representation and classification performance. The hierarchical architecture extracts complementary features using both networks, refined via projection and fully connected layers. Adaptive fusion dynamically weighs feature importance, while Squeeze-and-Excitation (SE) and Multi-Head Attention (MHA) enhance dependency learning and noise reduction. A residual connection ensures stable gradient flow and robust learning. For interpretability, SHAP and LIME provide global and local insights, fostering transparency and trust. With high accuracy and explainability, this model offers a reliable AI-assisted tool for colon cancer diagnosis.