| With the development of computer science and medical digitalization,the value of CT image target organ intelligent segmentation algorithm based on deep learning is growing in clinical diagnosis and surgery planning due to its advantages of efficiency,accuracy and ease of use.As the largest cavity in the human body,the abdominal cavity contains many important organs that are of high priority for research.However,the unstable and complex diversity of abdominal organs bring great challenges to the precise segmentation of multiple organs in the abdominal cavity.Therefore,this thesis proposes high-precision intelligent segmentation schemes for CT images of abdominal organs to address the above difficulties and the shortcomings of existing medical segmentation models.According to the different target organs of segmentation and their characteristics,the specific research of this thesis includes the following two parts:(1)Among the abdominal organs,the digestive organs are an important part of them.Due to their inherent characteristics,they generally have unstable morphology,and this main feature makes them different from other abdominal organs.Therefore,in this thesis,it is taken as the first subjects of research.Three important intestinal organs,namely,colon,small intestine and duodenum,are selected as representatives which are used to design the segmentation scheme.Firstly,based on the analysis of the CT images of colon,small intestine and duodenum,this thesis points out the highlights and difficulties for this task are twofold:one is the extraction of detailed features and multi-scale features,and the other one is the utilization of these two features.Secondly,for the first challenge,this thesis proposes to use High Resolution Network(HRNet)as the encoder in the segmentation model for the effective extraction of detailed and multiscale features.this thesis proposes to use High Resolution Network(HRNet)as the encoder in the segmentation model for the effective extraction of detailed features and multi-scale features.Thirdly,for the second challenge,based on channel attention mechanism,this thesis proposes the Scale Attention Module,which can greatly improve the model’s ability to exploit multi-scale features with the introduction of very few parameters and computational budget.In this thesis,we use this module as the core to build a segmentation model decoder and obtain the Scale Attention Network(SANet).Finally,the above two designs are validated to be effective and efficient for the high-precision intelligent segmentation of colon,small intestine and duodenum.SANet can obtain more precise segmentation results than the current leading medical segmentation models with the same number of parameters.The Dice Similarity Coefficient of the colon,small intestine,and duodenum reached 84.06%,76.79%,and 61.68%,respectively.In particular,SANet achieves far better performance than the comparative model for the small target duodenum.(2)Among the abdominal organs,besides the intestinal organs with unstable morphology,there are also a large number of important organs with relatively stable morphology.In this thesis,they are taken as the second subkects of research.Eight organs,including aorta,gallbladder,left kidney,right kidney,liver,pancreas,spleen and stomach are selected as representatives which are used to design the segmentation scheme.Firstly,based on the analysis of the prediction results of the existing medical segmentation models,this thesis points out the highlights and difficulties for this task are twofold:one is the analysis of the consistent feature of the difficult regions with diversity and complexity,the other one is the effective use of this consistent feature.Whereas previous work has tended to deal with difficult regions with certain specific characteristics due to their complexity and diversity,this thesis takes a more unified view of these difficult regions.By comparing the confidence heat map of the model prediction results and its annotation,this thesis points out that the consistent feature of the difficult regions is that the difficult regions are usually the regions with low confidence.Based on the above analysis,this thesis proposes a strategy for recalibrate the encoder output features using the previous confidence heat map,and introduces a recurrent structure to solve the non-causality dilemma,so as to construct a Recurrent Recalibration Network(R2Net)Finally,the experimental results on the Synapse abdominal organ segmentation dataset verified that the recurrent recalibration strategy could improve the performance of the existing medical segmentation model to a certain extent without introducing additional computation and parameters.The Dice Similarity Coefficient of the aorta,gallbladder,left kidney,right kidney,liver,pancreas,spleen and stomach reached 87.31%,71.69%,85.16%,82.26%,94.75%,66.97%,91.16%and 84.35%,respectively. |
