SAM2U-Net:基于SAM2与U-Net的医学图像分割模型  

SAM2U-Net:A U-Shaped Medical Image Segmentation Model Based on SAM2 and U-Net

作　　者：候俊伟 刘磊[1,2] Junwei Hou;Lei Liu(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai;School of Management,University of Shanghai for Science and Technology,Shanghai)

机构地区：[1]上海理工大学,光电信息与计算机工程学院,上海 [2]上海理工大学,管理学院,上海

出　　处：《建模与仿真》2025年第3期337-347,共11页Modeling and Simulation

摘　　要：医学图像分割在医学诊断中起着关键作用。尽管新兴视觉模型在各种医学分割任务中表现优异,但大多仅针对特定任务设计,缺乏普适性。本研究提出了一种新型SAM2学习模型,旨在实现通用医学图像分割。该模型基于U型架构,创新性地将SAM2的Hiera骨干网络与CNN模块并行结合,通过多尺度特征提取机制增强分割精度。在甲状腺结节诊断、结肠镜息肉分割等六个数据集上的实验表明,本模型在Dice系数和IoU上平均提高了1.46%,优于现有方法。结果证实,该模型能有效提取医学图像的病理特征,实现准确的区域分割,为广泛的临床诊断任务提供支持。Medical image segmentation plays a crucial role in medical diagnosis.Although emerging visual models perform excellently in various medical segmentation tasks,most are designed for specific tasks and lack universality.This study proposes a novel SAM2 learning model aimed at achieving general medical image segmentation.The model is based on a U-shaped architecture and innova-tively combines the Hiera backbone network of SAM2 with CNN modules in parallel,enhancing seg-mentation accuracy through a multi-scale feature extraction mechanism.Experiments on six da-tasets,including thyroid nodule diagnosis and colonoscopic polyp segmentation,demonstrate that this model improves the average Dice coefficient and IoU by 1.45%compared to existing methods.The results confirm that the model can effectively extract pathological features from medical im-ages,achieving accurate regional segmentation and supporting a wide range of clinical diagnostic tasks.

关 键 词：甲状腺结节 息肉分割 深度学习 超声图像 分割算法 医学图像 U型架构 Transformer模块 并行网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]