Transformers and LLMs as the New Benchmark in Early Cancer Detection

Abstract

The study explores the transformative capabilities of Transformers and Large Language Models (LLMs) in the early detection of Acute Lymphoblastic Leukaemia (ALL). The researchers benchmark Vision Transformers with Deformable Attention (DAT) and Hierarchical Vision Transformers (Swin) against established Convolutional Neural Networks (CNNs) like ResNet-50 and VGG-16. The findings reveal that transformer models exhibit remarkable accuracy in identifying ALL from original images, demonstrating efficiency in image analysis without necessitating labour-intensive segmentation. A thorough bias analysis is conducted to ensure the robustness and fairness of the models. The promising performance of the transformer models indicates a trajectory towards surpassing CNNs in cancer detection, setting new standards for accuracy. In addition, the study explores the capabilities of LLMs in revolutionising early cancer detection and providing comprehensive support to ALL patients. These models assist in symptom analysis, offer preliminary assessments, and guide individuals seeking information, contributing to a more accessible and informed healthcare journey. The integration of these advanced AI technologies holds the potential to enhance early detection, improve patient outcomes, and reduce healthcare disparities, marking a significant advancement in the fight against ALL.