Classifying lung and colon cancer from histopathological images remains a significant challenge due to the high degree of intra-class feature similarity and complex tissue morphology, particularly in lung cancer cases. While convolutional neural networks (CNNs) have demonstrated strong spatial feature extraction capabilities, they cannot inherently model long-range dependencies and global contextual relationships. Although attention-based methods partially address these limitations, they often suffer from overfitting, limited generalization across heterogeneous datasets, and insufficient interpretability for clinical adoption. To address these challenges, this study presents a Multi-Head Attention-Based Convolutional Neural Network (MHAB-CNN) ensemble framework that captures localized and global feature interactions critical for robust cancer classification.

A k-fold cross-validation strategy is adopted to train multiple MHAB-CNN models, from which the empirically top-performing ones are selected and aggregated to form a compact ensemble. This approach improves robustness, reduces overfitting, and ensures computational efficiency. Grad-CAM-based visualizations interpret the discriminative regions influencing the model’s predictions.

Experimental evaluation on the LC25000 dataset demonstrates that the proposed framework achieves an average validation accuracy of 99.84% across folds. Furthermore, the E3 ensemble configuration, comprising models M1, M6, and M9, achieves the highest classification score on the held-out test set.

The proposed MHAB-CNN ensemble framework effectively captures localized and global feature interactions critical for robust lung and colon cancer classification, while improving robustness, reducing overfitting, and enhancing interpretability for potential clinical adoption.