Histopathology-based diagnosis of lung and colorectal cancer is labor-intensive and subject to observer variability, motivating accurate and interpretable computer-aided diagnosis. In this paper, we propose a lightweight Histopathology Attention Network (HistoAttn-Net) for five-class lung and colon tissue classification on the LC25000 dataset. The model uses a compact convolutional backbone to extract local texture patterns and a multi-head self-attention block to capture long-range tissue context. To further improve robustness, we adopt a performance-aware ensemble strategy that selects the top-performing models from 10-fold cross-validation and aggregates their predictions via mean probability and majority voting. Explainability is provided through Grad-CAM applied at both the single-model and ensemble levels, yielding class-specific heatmaps that highlight clinically meaningful regions. On the held-out test set, the best single HistoAttn-Net achieves 99.6% accuracy, 99.5% macro Fl-score, and Cohen’s k of 0.996, while the top 3 ensemble attains 100% accuracy, macro Fl-score, and k, with macro ROC-AUC effectively equal to 1.0. These results indicate that the proposed framework offers an attractive balance of accuracy, efficiency, and interpretability for histopathology-based cancer diagnosis.