Authors: Sakshi Shukla, Surjya Narayan Sahoo
Abstract: Lung diseases continue to be one of the top causes of deaths worldwide, which means that there is a need for efficient and fast diagnosis procedures. In this paper, we provide an overview of recent research in automated detection of lung diseases from chest X-rays with Vision Transformer (ViT). Traditional deep learning solutions such as Convolutional Neural Networks (CNNs) have proven themselves to be very effective. However, they suffer from the inability to model the crucial dependencies between distant spatial elements. Vision Transformers represent a major breakthrough in deep learning through the use of self-attention mechanism to model global relationships between the different parts of the image. This paper provides an overview of recent developments in the area of ViTs, covering such topics as hybrid CNN+ViT architectures, dual-stream attention fusion, and gaze guided architectures. Comparative analysis of the results shows that hybrid approaches, where global relationships are combined with traditional local feature extraction methods, produce the best results with accuracy rates above 98%.
DOI: http://doi.org/10.5281/zenodo.21349712
