Research On Pancreas Automatic Segmentation Algorithm For CT Image

Pancreas segmentation aims to automatically mark the pancreas region from the abdominal CT image.Recently,pancreas segmentation has received increasing attention in the fields of computer-aided diagnosis and computer vision,and has become one of the hot research directions in the medical and computer science.As a precondition of computer-aided diagnosis technology,the automatic pancreas segmentation plays a vital role in locating tumors and diseased areas,observing the growth or reduction of tumor,providing surgical plans,and studying the anatomy of the pancreas.Despite the thrilling success of deep convolutional neural networks(DCNNs)in automatic pancreas segmentation,these methods still have some limitations.Aiming at the problems that pancreas appears in very small region and has complex anatomical structure as well as ambiguous boundary,this thesis proposes two automatic pancreas segmentation algorithms and conducts comprehensive experiments on the available benchmark datasets to verify the effectiveness of the proposed algorithms.The details of this thesis are as follows:First,to solve the problem of inaccurate localization of the existing coarse-to-fine pancreas segmentation algorithms,this thesis proposes an automatic pancreas segmentation algorithm based on reinforcement box-optimization-agent learning.We formulate the problem of pancreas segmentation as a Markov decision process,which consists of three sub-networks: a coarse localization network,a policy optimization network,and a segmentation execution network.In order to optimize the pancreas region obtained by the coarse localization network,this algorithm proposes a localization optimization strategy based on a two-branch reinforcement learning model.The strategy uses the way of deep Q-learning to make the box-optimization-agent learn to take appropriate actions under different states to obtain optimal pancreas target-context box pairs.The results on the NIH pancreas segmentation dataset show that the localization optimization strategy proposed in this thesis can obtain more accurate pancreas area,and then make segmentation execution network obtain more accurate pancreas segmentation results.Secondly,an automatic pancreas segmentation algorithm based on lightweight DCNN modules and spatial prior propagation is proposed to solve the following problems.The existing algorithms with many parameters are easy to cause overfitting,the computational complexity of these methods is heavy,and the current coarse-to-fine pancreas segmentation algorithms do not take into account the differences of the tasks between the two stages.In this thesis,we decouple the challenging small organ segmentation problem into a localization problem and a segmentation problem.Then we implement the two problems by designing two lightweight sub-networks that are function-specific: the localization sub-network and the segmentation sub-network.For the localization sub-network,we simplify the baseline network and introduce an inverted residual block to build a lightweight network with fewer parameters.For the segmentation sub-network,based on the localization sub-network,a scale-transferable feature fusion module is introduced to fuse the high-level and low-level features,which provides more powerful features for fine pancreas segmentation.On the one hand,this significantly reduces the computational complexity of the DCNN model for pancreas segmentation.On the other hand,with less trainable parameters,the proposed network would also alleviate the over-fitting problem as the annotated CT Scans acquirable for the training process are scarce.For obtaining more accurate pancreas regions,we also explore useful spatial priors and apply them in both the localization sub-network and the segmentation sub-network through the prior propagation module.Comprehensive experiments on the NIH dataset and the MSD dataset are conducted to evaluate the proposed approach.The experimental results demonstrate that this approach can effectively reduce the computational cost of the baseline model and can achieve the state-of-the-art performance when compared to the existing methods.