Towards Clinical Deployment: A Deployment-Oriented Lightweight Transformer for Low-Latency Medical Image Segmentation

Abstract

This paper presents a comprehensive investigation into the system design of lightweight Transformer architectures specifically tailored for medical image segmentation tasks. The standard Transformer model, while demonstrating remarkable performance in various domains, suffers from excessive parameters and high computational complexity when applied to medical imaging, which often involves high-resolution volumetric data. To address these challenges, we propose a series of lightweight improvements including: (1) a sparse attention mechanism that reduces computational burden by focusing on relevant regions of the image, (2) a modular design approach that enables flexible configuration of network components based on task requirements, and (3) parameter sharing and pruning techniques that eliminate redundant connections while maintaining model accuracy. The proposed system demonstrates significant advantages in clinical applications, particularly in real-time surgical navigation and telemedicine scenarios. By efficiently operating on resource-constrained devices such as portable ultrasound machines and mobile diagnostic platforms, the system enables precise medical image analysis with minimal latency. This technological advancement provides crucial technical support for the development of precision medicine and inclusive healthcare, offering potential solutions for resource-limited settings and remote healthcare delivery.