Abstract

This thesis evaluates the effectiveness of Vision Transformers (ViT) and Swin Transformers for breast cancer classification, highlighting their advantages over traditional Convolutional Neural Networks (CNNs) in processing cytological images. Amid the critical need for better breast cancer diagnostics, these transformer-based models emerge as a promising solution, adept at capturing complex spatial and contextual data in medical images.

The research methodology involved collecting and preprocessing a dataset of cytological and histopathological breast cancer images. The performance of the vision transformers was assessed using metrics such as accuracy, precision, recall, and AUC-ROC, and compared against established CNN architectures. The results demonstrate that vision transformers excel at extracting complex patterns from images, significantly outperforming current methods. Specifically, the study reports a 3.06% improvement in classification accuracy over traditional approaches, achieving 95.01% accuracy on test sets and perfect accuracy in validation.

The thesis underscores the potential of ViT and Swin models to advance early detection and diagnosis of breast cancer. Their success in the study suggests a transformative shift towards utilizing advanced deep learning architectures in medical image analysis. This approach not only enhances diagnostic accuracy but also offers a data-efficient solution to the challenges of breast cancer classification. The findings advocate for further exploration of transformer-based models, which could redefine the standards of computer-aided diagnosis and significantly impact the field of cancer classification.