Study On Automatic Image Segmentation In Tumor Surgical Navigation

Brain tumor segmentation is a crucial part in surgical navigation.The segmentation of brain tumor based on MRI(Magnetic Resonance Images)before surgery,doctors can obtain the location and size of tumors,and formulate related plans of surgery and treatment.Manual segmentation of brain tumors requires professional knowledge and costs huge amount of time,so automatic segmentation for MR images is necessary for clinical applications.Due to the complexity of brain tumor tissues and structures,and the problem of inherent imbalanced classes of tumors,automatic segmentation of brain tumors is still a big challenge.In view of the above problems,this paper proposed an automatic brain tumor segmentation model based deep learning.At the same time,we designed an auxiliary analysis system equipped with the algorithm that can provide objective auxiliary diagnosis for doctors.The main research works of this thesis are as follows:1.Build a 3D brain tumor segmentation model based on the auxiliary module and conditional random field.First,the auxiliary module was established.And the ROI(Region of Interest)extraction part and the adaptive cropping part in the auxiliary module were used to increase the attention to the segmentation targets.At the same time,a 3D segmentation network based on long and short connections was designed.Additional short skip connections were added in 3D network to compensate the missing information caused by the convolution.With the long skip connections,detailed features,including semantic information and spatial information,can be captured.And the weighted Dice loss of 3D network was developed to improve the attention of tumor sub-regions during the segmentation process,which can solve the problem of inherent imbalanced classes of tumors.At last,an optimization strategy based on conditional random field and the post-processing method was proposed to optimize the output of neural network.2.Evaluate the segmentation model using the BRATS dataset.On the one hand,through comparative experiments,we proved the effectiveness of the method using the BRATS2018 dataset.On the other hand,we also used the BRATS2015 dataset for online verification.the segmentation results that we obtained were compared and analyzed with the current advanced methods to verify the implementation level of the proposed method.3.Based on the proposed segmentation method,we designed an assisted diagnosis system for brain tumors.Considering the needs of doctors in actual work,the system provided the functions of storage,query,segmentation,visualization and download for brain tumor images.Using B/S architecture,the system was established by Flask framework to provide objective segmentation results and visual presentation for doctors.