动态蛇形卷积和图卷积注意力的氟骨症X光图像分割  

Dynamic snake convolution and graph convolution attention for skeletal fluorosis X-ray image segmentation

作　　者：黄蕾 吴云 LEI Huang;YUN Wu(State Key Laboratory of Public Big Data,Guiyang 550025,China;College of Computer Science and Technology,Guiyang 550025,China)

机构地区：[1]公共大数据国家重点实验室,贵州贵阳550025 [2]贵州大学计算机科学与技术学院,贵州贵阳550025

出　　处：《光学技术》2024年第6期745-753,共9页Optical Technique

基　　金：国家自然科学基金资助项目(62266011);贵州省科技计划项目(黔科合基础-ZK[2022]一般119)。

摘　　要：针对氟骨症X光图像中病灶区域轮廓不规则,边界模糊的问题,提出一种结合动态蛇形卷积、图卷积注意力的氟骨症X光图像分割方法。在编码器部分,提出全域动态蛇形卷积模块(ODSCM)来补充由于单一卷积所缺乏多尺度特征感知能力,从而减轻卷积过程中的特征丢失现象;在解码器部分,提出图卷积注意力模块(GCA)捕获长距离像素依赖,结合通道注意力聚焦关键病灶区域。为了提高性能并减少参数量,引入幻影卷积模块(GCB),并在U-Net解码结构的基础上去除了与前两层特征的跳跃连接操作,以提高模型的泛化性能。在氟骨症X光图像分割数据集上进行验证,所提方法的DSC、IoU、Accuracy和Recall分别为0.77、0.64、0.85和0.75。实验结果表明,所提方法优于目前主流的医学图像分割方法,特别在氟骨症细小骨质增生病灶上表现出色。To address irregular contours and border ambiguity in skeletal fluorosis X-ray images,A segmentation method that combines dynamic snake convolution and graph convolution attention is proposed.The omni-dimensional dynamic snake convolution module is introduced in the encoder to compensate for the lack of multi-scale feature perception in single convolutions,reducing feature loss.The graph convolution attention module in the decoder captures long-range dependencies and focuses on key lesion areas through channel attention.Additionally,the ghost convolution block enhances performance and reduces parameters.By eliminating the skip connection with the first two layers of features in the U-Net decoding structure,generalization performance improves.Validation on a skeletal fluorosis X-ray image segmentation dataset shows DSC 0.77,IoU 0.64,Accuracy 0.85,and Recall 0.75.Experimental results demonstrate the method's superiority in segmenting skeletal fluorosis small osteogenic lesions compared to current mainstream medical image segmentation methods.

关 键 词：动态蛇形卷积 图卷积注意力 氟骨症 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]