改进Segformer的前列腺超声图像语义分割算法  

Prostate ultrasound image semantic segmentation algorithm based on improved Segformer

作　　者：石勇涛[1,2] 柳迪 高超 杜威 邱康齐 SHI Yongtao;LIU Di;GAO Chao;DU Wei;QIU Kangqi(College of Computer and Information Technology,China Three Gorges University,Yichang 443002,China;Hubei Key Laboratory of Intelligent Vision Monitoring for Hydroelectric Engineering,China Three Gorges University,Yichang 443002,China)

机构地区：[1]三峡大学计算机与信息学院,湖北宜昌443002 [2]三峡大学湖北省水电工程智能视觉监测重点实验室,湖北宜昌443002

出　　处：《现代电子技术》2024年第15期65-72,共8页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(61871258):面向大坝变形监测的视频微小运动放大与四维视觉增强;湖北省中央引导地方科技发展专项(2019ZYYD007)。

摘　　要：前列腺超声图像在临床中的准确分割对后续诊断具有重要影响。因此,通过深度学习辅助实现前列腺边界的快速、准确分割非常必要。为此,文中提出了一种改进的前列腺分割网络(DA-Segformer)。利用Transformer、深监督和注意力机制,快速准确地分割前列腺超声图像。引入MAG模块提高网络对特征图和像素关联性的理解能力,以及对前景像素的敏感度。采用深监督策略,在解码过程中引入损失函数,优化梯度传播,增强网络对关键特征的学习表征能力。实验结果显示,在前列腺超声图像数据集上,DA-Segformer模型的mIoU、Dice系数、准确率和召回率等指标均优于其他主流语义分割模型。该方法有效解决了前列腺超声图像手工分割的难题,为临床诊断提供了有价值的计算机辅助工具。Accurate segmentation of prostate ultrasound images in clinical settings plays a significant role in subsequent diagnosis.Therefore,it is essential to rapidly and accurately segment the prostate boundary with deep learning assistance.To this end,a novel prostate segmentation network named DA-Segformer is proposed.In this network,the Transformer,deep supervision and attention mechanism are utilized to segment prostate ultrasound images rapidly and accurately.Additionally,the MAG module is introduced to enhance the network's understanding of feature maps and pixel correlations,so as to improve its sensitivity to foreground pixels.A deep supervision strategy is employed.A loss function is introduced into the decoding process to optimize gradient propagation,so as to enhance the network's ability to learn and represent the key features.Experimental results demonstrate that the mIoU(mean intersection over union),Dice coefficient,accuracy rate and recall rate of the DA Segformer model on the prostate ultrasound image dataset are superior to those of the other mainstream semantic segmentation models.The proposed method effectively addresses the challenge of manual segmentation of prostate ultrasound images,and provides valuable computer-aided tools for clinical diagnosis.

关 键 词：医学图像分割 超声图像分割 TRANSFORMER 门控注意力 深监督 扩张卷积 梯度下降 多尺度特征 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]