摘要
针对口腔移植骨区域CBCT图像边界模糊、结构差异性大、手动分割耗时等问题,提出了一种基于端到端的神经网络分割算法。该算法以UNet++为基础分割网络,创新性地设计密集的空洞多尺度模块提取不同大小的移植骨区域的特征,并在特征融合前嵌入坐标注意力块、在分割网络的末端融入了尺度注意力模块以提高重要区域的权重。基于上海市第九人民医院提供的CBCT图像数据集上进行验证与评估,与原始的UNet++相比,该算法的Dice系数、Jaccard系数、敏感度、精确率分别提升了4.92%、4.96%、7.64%和2.53%。
This paper proposes an end-to-end neural network segmentation model to handle the challenges of blurred boundaries of CBCT images in the bone graft region,large structural differences,and the time cost of manual segmentation.The proposed algorithm takes UNet++as the basic network,where an atrous multiscale module is innovatively designed to extract the features of different sizes of grafted bone regions,and embedded a coordinate attention block embedded feature fusion and scale attention added at the end of segmentation networks are utilized to increase the weight of the region of interest.Based on the verification and evaluation on the CBCT image data set provided by Shanghai Ninth People's Hospital,compared with the original UNet++,the algorithm's Dice coefficient,Jaccard coefficient,sensitivity,and accuracy have increased by 4.92%,4.96%,7.64%and 2.53%respectively.
作者
赵宇
徐常鹏
丁德锐
ZHAO Yu;XU Changpeng;DING Derui(School of Optical Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093)
出处
《计算机与数字工程》
2025年第6期1704-1710,共7页
Computer & Digital Engineering
基金
国家自然科学基金项目(编号:61973219)资助。
