基于改进Unet与动态阈值可变FCMSPCNN的医学图像分割

针对深度学习的医学图像分割模型训练时间长和精度不精的问题，提出结合动态阈值可变FCMSPCNN的多尺度上下文编解码结构和注意力机制的CoA Unet(Context Attention Unet)分割方法。首先，使用动态阈值可变的FCMSPCNN预分割出目标矩形区域并使用掩码遮盖背景部分；然后，更深层卷积块加入快捷连接交叉融合不同层次的特征，并通过注意力门突出对目标特征的学习；最后，在编解码器最底层加入改进的多尺度上下文提取器可以更好地提取目标特征信息。模型分别在LiTs和DRIVE数据集上进行验证，肝脏分割指标Miou、Aver＿HD、Aver＿Dice分别为0.890 5、6.369 9、0.947 7，视网膜血管分割指标分别为0.589 2、9.255 9、0.740 9。实验表明，预处理能缩短4.3%～20.33%的训练时间并提升2%～6%分割精度，与其他5种分割方法相比，CoA Unet能取得更好的整体分割性能。

Medical image segmentation using improved Unet combined with dynamic threshold changed FCMSPCNN

Abstract: Aiming at the long training time and low precision of medical image segmentation model based on deep learning, a segmentation method combining the multi-scale context encoding and decoding structure of dynamic threshold changed FCMSPCNN with context attention Unet (CoA Unet) is proposed. Dynamic threshold changed FCMSPCNN is used to pre-segment the target rectangle and mask the background. The deeper convolution block with shortcut connection integrates features at different levels, and highlights the learning of target features through the attention gate. Furthermore, an improved multi-scale context extractor is added to the bottom layer of the codec to better extract target feature information. The model is verified on LiTs and DRIVE datasets, respectively. The Miou, Aver_HD and Aver_Dice of the proposed method are 0.890 5, 6.369 9, 0.947 7 for liver segmentation, and 0.589 2, 9.255 9, 0.740 9 for retinal vascular segmentation. The experiment reveals that the preprocessing can shorten the training time by 4.30%-20.33% and improve the segmentation accuracy by 2%-6%. Compared with other 5 segmentation methods, CoA Unet can achieve better overall segmentation performance.