Cancer Identification in Bowel Tissue Images Using CNNs and Transformers

Cancer remains a leading cause of mortality worldwide, necessitating advances in diagnostic methodologies for early and accurate detection. This study presents a comprehensive comparison of traditional classifiers, Convolutional Neural Networks (CNNs), and vision transformers in distinguishing cancerous from non-cancerous bowel tissue images obtained from laboratory rats. Using a novel dataset of 805 images, we evaluated traditional classifiers with Local Binary Pattern (LBP) features, deep learning architectures (ResNet-50), Vision Transformer (ViT), and Data-Efficient Image Transformer (DeiT). ResNet-50 demonstrated remarkable performance (91.70% accuracy, 0.912 F1 score), while vision transformers, despite extensive optimization, achieved approximately 78% accuracy, revealing their limitations with smaller datasets. Most significantly, ensemble methods combining CNN and transformer architectures (specifically ResNet-50 and DeiT) yielded superior results (92.30% accuracy, 0.955 AUC), demonstrating how these fundamentally different architectural approaches can complement each other.