Accurate and timely diagnosis of colorectal cancer is critical for improving patient outcomes. While histopathological examination remains the gold standard, it is labor-intensive and prone to inter-observer variability. This study investigates the application of deep learning, specifically ResNet and Swin Transformer V2 architectures, for automated CRC subtype classification using hematoxylin and eosin stained histopathology images. We evaluate lightweight variants-ResNet-18, ResNet-34, ResNet-50, Swin v2-T Small, Swin v2-T Tiny, and Swin v2-T Tiny w8-for their ability to extract relevant visual features. To enhance model generalizability under limited data conditions, we apply data augmentation techniques including geometric and color transformations, elastic deformations, and random cropping. Model interpretability is addressed using Grad-CAM, enabling visualization of regions contributing most to predictions. Among the models, ResNet-34 achieved the best trade-off between accuracy and complexity, with overall, top-2, and top-3 classification accuracies of 91.10%, 99.11%, and 100.00%, respectively. Swin Transformer V2 models also showed competitive results, particularly in adenocarcinoma (up to 99.17%) and polyp (95.14%) detection, though at the cost of increased computational demands. These results suggest that moderate-depth architectures can effectively capture the morphological diversity of colorectal cancer subtypes. Our approach provides an interpretable, efficient deep learning-based diagnostic tool that can support pathologists by improving classification accuracy and consistency, ultimately contributing to more personalized and timely cancer care.