<![CDATA[The diagnosis of colorectal cancer (CRC) through histopathological images requires high accuracy to support appropriate clinical decisions. Although Convolutional Neural Networks (CNN) have become the gold standard in medical image analysis, the emergence of Vision Transformer (ViT) architecture offers a new paradigm based on global attention mechanisms (self-attention) that is claimed to be superior on large-scale datasets. However, the effectiveness of ViT-B/16 on medical datasets with limited sample sizes and high texture variation remains debatable. This study aims to comprehensively evaluate the performance of the ViT-B/16 architecture compared to ResNet-50 on the NCT-CRC-HE-100K histopathology dataset, which consists of 9 network classes. The performance of both models was tested using equivalent training scenarios. The evaluation was conducted multidimensionally, covering classification metrics (Accuracy, F1-Score), training stability, feature space separability (t-SNE), visual interpretability (Grad-CAM), and computational efficiency. The experimental results show that ResNet-50 significantly outperforms ViT-B/16 with a test accuracy of 93.24%, compared to ViT-B/16 which only achieves 57.11%. The t-SNE analysis revealed that ViT-B/16 failed to form well-separated feature clusters due to a lack of inductive bias to recognize local features such as cell membrane edges. Failure analysis shows that ViT-B/16 often misclassifies adipose cells as mucus and smooth muscle as tumors. In terms of efficiency, ResNet-50 is 5.8 times lighter in storage size and has lower inference latency. This study concludes that CNN-based architecture (ResNet-50) is still far superior, more stable, and more feasible for clinical implementation than ViT-B/16 in the context of medium-scale histopathological image classification.]]>
