融合CNN-Transformer的医学图像分割网络

Medical Image Segmentation Network Integrated with CNN-Transformer

主流卷积神经网络在医学图像分割中通常面临三个挑战。首先,常规卷积运算主要获取医学图像的局部特征,其在图像长程信息建模能力方面表现出局限性;其次,卷积神经网络中常规的下采样操作会导致医学影像特征图中的重要信息丢失,影响分割效果;最后,当卷积运算带来的问题得以解决时,如何将提取到的局部特征和全局特征充分融合。为解决上述问题,提出了一种融合CNN和Transformer的医学图像分割网络。该网络首先通过引入Transformer来解决卷积运算感受野固定的问题;其次使用Patch Embedding来解决下采样过程中重要信息丢失的问题;最后通过交替使用CNN和Transformer来解决局部特征和全局特征无法充分融合的问题。在ISIC2018和KiTS19数据集上的实验结果表明,提出的网络不仅能够捕捉更精细的轮廓弧度,并且有较强的抗干扰能力,具有较高的分割精度和鲁棒性。

Mainstream convolutional neural networks usually face three challenges in medical image segmentation.First,conventional convolutional operations mainly acquire local features of medical images,which show limitations in image long-range information modeling capabilities.Secondly,the conventional downsampling operation in convolutional neural networks will lead to the loss of important information in the feature map of medical images,affecting the segmentation effect.Finally,when the problems caused by convolutional operations are solved,how to fully integrate the extracted local features and global features.In order to solve the above problems,a medical image segmentation network integrating CNN and Transformer is proposed.The network first solves the problem of convolutional operation sensor field fixation by introducing Transformer.Second,Patch Embedding is used to solve the problem of important information loss during the downsampling process.Finally,the problem of insufficient fusion of local and global features is solved by alternating cnns and transformers.Experimental results on the ISIC2018 and KiTS19 datasets show that the proposed network is not only able to capture finer contour arcs,but also has strong anti-interference ability,with high segmentation accuracy and robustness.