摘要
针对医学图像分割中上下文信息联系匮乏和网络过深导致分割精度低的问题,提出了一种基于改进UNet的脑肿瘤图像分割算法。嵌套残差连接,组成一种深度监督网络模型,用密集跳跃连接替换UNet传统的连接方式,减小编码路径和解码路径特征图之间的语义差距。在残差模块中加入了注意力机制和软阈值化函数,有效降低图像噪声,减少了网络梯度的弥散程度。实验结果在全肿瘤区域(WT)Dice系数为0.846,肿瘤核心区域(TC)Dice系数为0.813,肿瘤增强区域(ET)Dice系数为0.804。验证结果表明,该方法能在复杂情况下对脑部肿瘤模糊边界进行精细分割,提高分割精度。
To solve the problem of low segmentation accuracy caused by lack of context information connection and excessive network depth in medical image segmentation,a brain tumor image segmentation algorithm based on improved UNet is proposed.A deeply supervised network model is formed by nested residual connections,and the traditional UNet connections are replaced by dense skip connections to reduce the semantic gap between encoding path and decoding path feature maps.The attention mechanism and soft thresholding function are added into the residual module,which can effectively reduce the image noise and reduce the diffusion degree of network gradient.The experimental results show that the Dice coefficient is 0.846 in the whole tumor region,0.813 in the tumor core region and 0.804 in the enhanced tumor region.It is proved that this method can accurately segment the fuzzy boundary of brain tumor in complex conditions,thus improving the segmentation accuracy.
作者
王莹
朱家明
徐婷宜
宋枭
WANG Ying;ZHU Jiaming;XU Tingyi;SONG Xiao(College of Information Engineering,Yangzhou University,Yangzhou 225009,China)
出处
《无线电工程》
北大核心
2022年第9期1566-1573,共8页
Radio Engineering
基金
国家自然科学基金。
