A 3D Brain Tumor Segmentation Algorithm Based On Optimal Transportation

Lesion segmentation in medical imaging refers to identifying and segmenting the semantic information of lesions in organs effectively by proper image segmentation algorithms.Efficient medical image segmentation methods can provide effective assistance for doctors in the diagnosis.The shape of original brain image is irregular.Therefore,it is necessary to extract the image information and combine it into data in tensor form as input data before deep-learning,which is a kind of popular method for image processing.The data in tensor form is used as input data.Nowadays,many methods are used in medical image preprocessing.However,how to extract the global information of brain from the original image effectively is still a hot research subject in academia.This thesis aims to study the three-dimensional brain tumor segmentation problem and its numerical algorithm based on the two-stage optimal transportation preprocessing method.This thesis mainly consists three parts.In the first part,the optimal transportation method is used to preprocess the brain data,aiming to transport a irregular brain image to another location without obvious distortion: The first step is to transform the brain into a unit sphere through a mass-preserving transportation mapping where the spherical boundary constraints are imposed to ensure the convergence of the algorithm.The second step is to transform the unit sphere into a unit cube by the inverse optimal mass transportation mapping.Two-step image transformation is introduced to construct a two-stage optimal transportation mapping in this part,which realizes the conversion from the brain with the dimension(240,240,240,4)to the cube with the dimension(128,128,128,4).In the second part,the 3D U-net algorithm is introduced to construct three corresponding neural networks for the overall tumor,the tumor core and the enhanced tumor respectively,in which the cube training data is input for model training.In the third part,the trained model is applied to predict the data in the test dataset,and then restore the the predicted cube data to the original brain shape by the inverse mapping of the two-stage optimal transportation mapping.Thus the result of brain tumor segmentation is obtained..For 484 brain images,the Dice training accuracy of the three tumors can reach 96.64%,94.91%,91.61% and the test accuracy can reach 90.18%,87.64%,81.81%.