Research On Pancreas Automatic Segmentation Algorithm In CT Scans Based On Attention Mechanism

Pancreas segmentation refers to marking the area of the pancreas from CT scans.As a pre-step of computer-aided diagnosis technology,it can be used for pancreas localization,disease detection,and surgical planning and arrangement.It can play a huge role in clinical application.Therefore,designing an automatic pancreas segmentation algorithm with good performance is of great significance in the field of medical image segmentation and computeraided diagnosis.With the rapid development of convolutional neural networks（CNN）,traditional pancreas segmentation algorithms based on handcrafted features have been gradually replaced by two-stage algorithms based on CNN.Although the automatic pancreas segmentation algorithm based on CNN has achieved great success,the existing algorithms still have some shortcomings.In this paper,we propose two automatic pancreas segmentation algorithms to address the differences between localization and segmentation in the two-stage approach and the problem that the shallow pancreas shape features are easily lost and conduct extensive experiments on two datasets to prove the effectiveness of the algorithms.The main research of this paper is as follows:Firstly,the current algorithms decompose pancreas segmentation into two subtasks（localization and segmentation）and use the same network structure to deal with different tasks,which is unreasonable.Because the generalization ability of the model is limited.To solve this problem,we propose a pancreas segmentation algorithm by two-view feature learning based on attention mechanism and multi-scale supervision.This algorithm consists of a location branch,a weight conversion model,and a segmentation branch.According to the differences between localization and segmentation,we use different attention modules and design specific network frameworks for the location and segmentation branches.In addition,this algorithm is supervised in three scales,each with its target,i.e.,a localization loss for improving the location branch’s accuracy for image-level pancreas localization,an auxiliary loss forcing the weight conversion module to learn global semantic information,and a segmentation loss for improving the model’s sensitivity for pixel-level pancreas segmentation.At the same time,the input of the segmentation branch is optimized in a circular manner,which eliminates the excessive dependence on the pancreas localization and avoids the negative impact of incorrect localization results.Secondly,we design a pancreas segmentation algorithm by feature propagation and fusion based on attention mechanism to solve the problem that the shape features of the pancreas are lost during convolution.The algorithm also divides the pancreas segmentation into two sub-tasks（localization and segmentation）and uses two sub-networks----detection network and segmentation network to solve these two sub-tasks respectively.In both subnetworks,the algorithm designs a progressive feature propagation with fusion of shallow features and deep features to maintain,propagate and facilitate the reuse of shallow shape features.In addition,the algorithm uses Attentional Feature Fusion（AFF）based on contextual information to eliminate the problems arising from the small size and large shape differences of the pancreas.Finally,the algorithm designs a simple and effective attentional module in the segmentation network for fine-grained segmentation of the pancreas.The two algorithms proposed in this paper have been extensively experimented and achieved excellent performance on the NIH pancreas segmentation dataset as well as the pancreas dataset in the Medical Segmentation Decathlon.Compared with the current mainstream automatic pancreas segmentation algorithms,the algorithms in this paper show significant improvements in several evaluation metrics.