基于对比学习增强的Lora微调超声影像分割模型

A Contrastive Learning-Enhanced Lora Fine-Tuned SAM-Med3D Model for Ultrasound Image Segmentation

摘要: 超声影像分析在现代医学中扮演着至关重要的角色，但精确分割是其面临的主要挑战之一。尽管现有的深度学习模型如SAM在自然图像上表现出色，但在医学图像分割上仍存在性能差距。本研究提出了一种基于对比学习增强的LoRA微调SAM-Med3D超声影像分割模型(USCL-Med3D)，旨在提高3D超声影像分割的精确度和效率。为此，设计了一种半监督伪标签数据集训练方法，通过自动化获取标注数据，降低了标注难度并保证了标注效果。同时，引入对比学习架构VCL-head，增强了模型对3D超声影像上下文信息的提取能力。此外，还对SAM-Med3D模型进行了LoRA微调，从而使模型具有更好的分割能力。实验结果表明，所提方法在3D超声数据集和一些公开的3D医疗影像数据集上取得了优异的分割效果。

Abstract: Ultrasound image analysis plays a critical role in modern medicine, but precise segmentation remains one of its major challenges. Although existing deep learning models like SAM perform well on natural images, there is still a performance gap in medical image segmentation. This study proposes a contrastive learning-enhanced LoRA fine-tuned SAM-Med3D ultrasound image segmentation model (USCL-Med3D) to improve the accuracy and efficiency of 3D ultrasound image segmentation. We designed a semi-supervised pseudo-label dataset training method to automatically obtain annotated data, reducing annotation difficulty while ensuring annotation quality. Additionally, a contrastive learning architecture was introduced to enhance the model’s ability to extract contextual information from 3D ultrasound images. Furthermore, we fine-tuned the SAM-Med3D model using LoRA, effectively incorporating the feature representation abilities of the 3D ultrasound dataset. Our method achieved excellent segmentation performance on the 3D ultrasound dataset and several publicly available 3D medical imaging datasets.