基于Adapter微调的轻量型SAM在超声图像分割中的研究

Research on Lightweight SAM Based on Adapter Fine-Tuning for Ultrasound Image Segmentation

摘要: 随着医学影像技术的快速发展，超声图像在疾病诊断和治疗中发挥着重要作用。然而，超声图像具有噪声大、对比度低和边界模糊等特点，给超声图像分割任务带来了巨大挑战。本文提出了一种基于Adapter微调的轻量超声图像分割模型LG-MedSeg，旨在解决计算资源受限和超声图像分割精度不足的问题。通过知识蒸馏技术将SAM-Med2d模型轻量化，并引入并行低秩分解Adapte (LRAdapter)模块，增强模型对超声图像特征的适应能力。实验结果表明，LG-MedSeg在胸腔积液和腹腔积液分割任务中表现出色，Dice系数达到90.85%，IoU达到84.5%，同时模型参数量仅为13.15 M，显著降低了计算复杂度，并在多个公开数据集上的实验验证了LRAdapter的有效性和通用性。本文的研究为医学图像分割提供了一种高效、轻量化的解决方案，特别适用于资源受限的边缘设备部署。

Abstract: With the rapid development of medical imaging technology, ultrasound images play a crucial role in disease diagnosis and treatment. However, ultrasound images are characterized by high noise, low contrast, and blurred boundaries, posing significant challenges to ultrasound image segmentation tasks. This paper proposes LG-MedSeg, a lightweight ultrasound image segmentation model based on Adapter fine-tuning, aiming to address the issues of insufficient computational resources and inadequate segmentation accuracy for ultrasound images. By applying knowledge distillation technology to streamline the SAM-Med2D model and introducing a parallel low-rank decomposition Adapter (LRAdapter) module, the model’s adaptability to ultrasound image features is enhanced. Experimental results demonstrate that LG-MedSeg achieves outstanding performance in thoracic effusion and abdominal effusion segmentation tasks, with a Dice coefficient of 90.85% and IoU of 84.5%, while maintaining only 13.15 million parameters. This significantly reduces computational complexity, and experiments on multiple public datasets validate the effectiveness and generalizability of the LRAdapter. Our research provides an efficient and lightweight solution for medical image segmentation, particularly suitable for deployment on resource-constrained edge devices.