基于双任务一致性的半监督深度学习医学图像分割方法

Semi-Supervised Medical Image Segmentation Method Based on Dual Task Consistency

【目的】医学图像分割是医学图像分析中的一个重要内容。现有的大部分图像分割算法都是基于监督学习,而实际应用中医学图像标签难以获取,大量标注需依赖领域专家,费时费力。因此,提出一种双任务一致性的半监督医学图像分割模型。【方法】该模型采用一个编码器,两个解码器的网络结构,其中编码器和一个解码器实现图像分割,与另一个解码器实现图像重建。无标签数据通过一致性分割与重建任务得到的两个不同图像背景计算损失并优化网络参数。同时在网络的编码器部分加入注意力模块以更好地获取分割区域的空间位置信息,并使用锐化操作增加无标签数据重建输出的置信度。【结果】在肝脏和细胞数据集上IOU分别为0.953 6和0.821 0,相较于U-Net提高了1.5%和4.82%;在眼底血管数据集上,SP值为0.983 0,与第二名相比提升了0.18%。【结论】本文模型与有监督方法和半监督方法相比,在医学图像分割的有效性和泛化性上有一定的性能提升,能有效解决数据集数量少、小病灶分割难度大的问题。

[Purposes]Medical image segmentation is an important content in medical image analysis. Most of the existing image segmentation algorithms are based on supervised learning so far, but the actual application of medical image label is difficult to obtain, numerous labeling need to rely on domain experts, which would cost time and labor. To solve the problems, a dual task consistent semi-supervised medical image segmentation model is proposed. [Methods]The model used a network structure of one encoder and two decoders, where one decoder did image segmentation and the other decoder did image reconstruction. The loss of unlabeled data was calculated and network parameters were optimized through two different image backgrounds obtained by consistent segmentation and reconstruction task. Simultaneously, the attention block is added to the network to better obtain the spatial location information of the segmented region, and sharpening operation is used to increase the confidence of the reconstructed network image. [Findings]The IOU of liver and cell were 0.953 6 and 0.8210 and the SP value of retinal blood vessel was 0.983 0. [Conclusions]Compared with the latest supervised methods in medical image segmentation, the proposed model also has a certain performance improvement in terms of effectiveness and generalization.