Non-Hodgkin Lymphoma (NHL) encompasses a heterogeneous group of hematologic malignancies, presenting considerable diagnostic complexity due to diverse morphological subtypes. Accurate and automated classification of NHL is essential for effective treatment planning and prognosis. This study proposes a deep learning-based diagnostic framework leveraging a pre-trained InceptionV3 convolutional neural network optimized using the Dung Beetle Optimizer (DBO), a nature-inspired metaheuristic algorithm. The optimized InceptionV3-DBO model is fine-tuned to enhance feature extraction and hyperparameter tuning for improved classification of three major NHL subtypes: Chronic Lymphocytic Leukemia (CLL), Follicular Lymphoma (FL), and Mantle Cell Lymphoma (MCL). The histopathological image dataset, curated from publicly available sources, was subjected to stain normalization and augmentation to address staining variability and class imbalance. Experimental evaluation using k-fold cross-validation demonstrates that the proposed InceptionV3-DBO outperforms conventional CNN architectures (CNN, VGG16, DenseNet121, ResNet50) and a recent Vision Transformer (ViT) baseline in terms of accuracy, sensitivity, specificity, F1-score, MCC, and Kappa coefficient. Moreover, Grad-CAM visualization is integrated to provide interpretable heatmaps highlighting discriminative regions, supporting clinical transparency and trust. The model achieves an MCC of 95.9% and competitive accuracy, confirming its potential for real-world deployment in computer-aided lymphoma diagnostics.