Research On Soft Tissue Organ Auto-Segmentation Based On Statistical Prior

Accurate segmentation of soft tissue organ medical images is the cornerstone of follow-up clinical diagnosis and scientific researches.However,due to the uneven distribution of gray scales,large differences among different biological individuals,noise and artifacts in soft tissue organ medical images,it is difficult to obtain accurate segmentation results using common image segmentation methods.In order to overcome this difficulty,this thesis conducted an in-depth study on the soft tissue organ auto-segmentation method based on statistical priors,and used the statistical prior information obtained from training samples to guide soft tissue organ auto-segmentation,improving the accuracy of soft tissue organ segmentation.In this thesis,two different methods based on statistical priors for soft tissue organ auto-segmentation were mainly studied,and the experiments were conduted separately.The research work of this thesis can be summarized as follows:(1)Aiming at the complexity of soft tissue organ segmentation task,according to the idea of staged and multi-task,this thesis proposed a soft tissue organ segmentation method based on landmark registration and a 3D fully convolutional network,which divided the image segmentation task into multiple step-by-step subtasks.Firstly,landmarks with uniform constraints were defined in training images,and were used to train a regression forest model to detect the landmarks in testing images.Then,a shallow neural network was trained to guide the mapping of targets from training images to testing images based on the statistical prior information provided by these landmarks.Finally,a 3D fully convolution network was used to fine-tune the contour of target,obtaining the final segmentation result.This method was applied in the segmentation of heart target in human CTA images,and the accuracy,Dice coefficient and Hausdorff distance of segmentation result were 96.25%,93.98% and 2.12 voxels,respectively,which were superior to those of all comparison methods.The experimental results show the superiority of the multi-task integration segmentation method proposed in this thesis in complex soft tissue organ segmentation,which has great potential for clinical application.(2)In order to evaluate the effectiveness of the soft tissue organ segmentation method better,from the perspective of the subsequent applications of segmentation process,an evaluation method based on small animal fluorescence molecular tomography(FMT)reconstruction accuracy was proposed in this thesis,used to evaluate active shape model(ASM)based segmentation method that incorporated lots of statistical prior information.Firstly,a shape model was established based on the statistical prior information obtained in training images,then the model deformed and was used in testing image segmentation based on the gray search strategy,and finally the fluorophore position in small animal was obtained by FMT reconstruction.By performing kidney segmentation and fluorophore reconstruction experiments on mouse micro-CT images,the Dice coefficient,Hausdorff distance and reconstruction distance error of ASM-based segmentation method reached90%,0.30 mm and 1.75 mm,respectively,which were not only superior to those of comparison method,and the reconstruction distance error was also closed to that(1.70mm)of manual segmentation method(gold standard).The experimental results show the role of statistical prior in soft tissue organ segmentation,and verify the feasibility of using FMT reconstruction accuracy to evaluate the segmentation methods.In summary,this thesis mainly studies two soft tissue organ segmentation methods based on statistical prior.The research were conducted from the perspective of the segmentation method itself and the evaluation of segmentation methods,respectively,showing the superiority of soft tissue organ segmentation methods based on statistical prior and the feasibility of using FMT reconstruction accuracy to evaluate segmentation methods.