Towards efficient diagnostics: refining vision transformers for medical image multi-label classification

Medical imaging, including the use of chest X-rays, is an important tool for modern healthcare, enabling early and accurate disease diagnosis, facilitating timely interventions to mitigate health issues. By capturing images of critical internal organs like the lungs and heart, X-rays enable doctors to make informed diagnoses and treatment decisions, especially concerning respiratory and cardiac conditions. The importance of early and accurate disease diagnosis, particularly with multiple pathologies, is paramount, as it greatly impacts patient outcomes by enabling timely and specific treatments. Recently, multi-label classification has become increasingly important in medical imaging, since several pathologies can be present within a single X-ray. While traditional convolutional neural networks (CNNs) have played a pivotal role in enhancing the accuracy of X-ray diagnoses, the expanding complexity of multi-label imaging demands more sophisticated methods. Vision Transformers (ViTs) have emerged as a promising approach in medical image classification, showcasing their ability to effectively process X-ray images and identify pathologies within them. While traditional ViTs perform well, they have significant drawbacks. Most ViT models utilize a large number of parameters, often ranging from millions to billions of parameters. Such parameter-intensive designs, while powerful, are computationally heavy. This not only increases the resource requirements, but also raises concerns about their feasibility and scalability in real-world, time-sensitive healthcare settings. We propose a novel Vision Transformer architecture aimed at effectively classifying multi-label X-ray images while significantly enhancing the efficiency of ViT-based multi-label medical image classification methods. By optimizing model architectures and exploring techniques for parameter reduction, we seek to develop more streamlined and resource-efficient approaches without completely sacrificing the efficacy of these methods. Our work endeavors to bridge the gap between cutting-edge technology and practical healthcare applications, promising a more efficient and accessible future for medical image analysis.