基于循环显著性校准网络的胰腺分割方法

胰腺的准确分割对于胰腺癌的识别和分析至关重要. 研究者提出通过第一阶段粗分割掩码的位置信息缩小第二阶段细分割网络输入的由粗到细分割方法, 尽管极大地提升了分割精度, 但是在胰腺分割过程中对于上下文信息的利用却存在以下两个问题: 1) 粗分割和细分割阶段分开训练, 细分割阶段缺少粗分割阶段分割掩码信息, 抑制了阶段间上下文信息的流动, 导致部分细分割阶段结果无法比粗分割阶段更准确; 2) 粗分割和细分割阶段单批次相邻预测分割掩码之间缺少信息互监督, 丢失切片上下文信息, 增加了误分割风险. 针对上述问题, 提出了一种基于循环显著性校准网络的胰腺分割方法. 通过循环使用前一阶段输出的胰腺分割掩码作为当前阶段输入的空间权重, 进行两阶段联合训练, 实现阶段间上下文信息的有效利用; 提出卷积自注意力校准模块进行胰腺预测分割掩码切片上下文信息跨顺序互监督, 显著改善了相邻切片误分割现象. 提出的方法在公开的数据集上进行了验证, 实验结果表明其改善误分割结果的同时提升了平均分割精度.

Pancreas Segmentation Based on Recurrent Saliency Calibration Network

Accurate segmentation of the pancreas is very important for the identification and analysis of pancreatic cancer. The researchers proposed a coarse-to-fine segmentation method to reduce the input of the second-stage fine segmentation network through the position information of the first-stage coarse segmentation mask. Although the segmentation accuracy is greatly improved, however, the use of context information in the pancreas segmentation process has the following two problems: 1) The coarse segmentation and fine segmentation stages are trained individually, and the fine segmentation lacks the predicted mask information of the coarse segmentation, which suppresses the flow of context information between stages, resulting in part of the fine segmentation that cannot be more accurate than the coarse segmentation; 2) In the coarse and fine segmentation stage, there is a lack of mutual supervision information between the adjacent predicted masks of a single batch, which leads to the loss of inter-slice context information and increases the risk of false segmentation. To solve the above problems, a pancreas segmentation method based on the recurrent saliency calibration network is proposed. By recurrently using the previous stage output segmentation mask as the spatial weight of the current stage input and performing joint training, the context information between stages is effectively used. Besides, a convolutional self-attention calibration module is suggested, which performs cross-sequence supervision of inter-slice context information and significantly improves the false segmentation. The proposed method is verified on the public datasets, and the experimental results show that it improves the average segmentation accuracy while improving the results of false segmentation.