A Two-Phase 3D Brain Tumor Image Segmentation Method

Due to the complex structure of the human brain,lesions arise in a variety of tissues and areas.In terms of technical di culties and diagnostic e(?)ectiveness,segmenting lesions in MRI brain medical pictures presents a challenge to doctors and professionals.As a result,the design of automatic segmentation technologies for the brain tumor pictures is becoming increasingly significant.To address this issue,this paper will combine the optimal transmission algorithm with quality-preserving and a 3D Res UNet network to create a two-phase deep learning model for realizing the whole tumor(WT),tumor core(TC),and enhanced tumor(ET)sub-regions in brain tumor images for accurate segmentation and prediction.The following three sections include the particular research content:First,this paper proposes the quality-preserving optimal transport(VOMT)algorithm,which is utilized to preprocess the data from the 3D MRI brain medical pictures,converting irregular three-dimensional information into a regular cubes,so that the data matches the deep learning model’s input format.Second,an e cient two-stage deep learning model 2P-Res UNet-OMT is built using the VOMT algorithm and the 3D Res UNet network.The VOMT method is used in the first stage of the model to preprocess the brain imaging data before it is imported into the 3D Res UNet model for the predicrion of the WT region.In the second stage,this region is enlarged outward by two voxels,and the mesh of the expanded region is refined subsequently.The brain is again mapped into a cube using VOMT and then placed into the 3D Res UNet model for training.Finally,ensemble voting post-processing is utilized to improve segmentation accuracy.Finally,using the 1251 dataset o(?)ered by the Bra TS 2021 competition,we run associated numerical experiments.We randomly selected 1000 and 251 brain samples from the Bra TS 2021 training dataset for training and validation respectively.Under validation computed by numerical experiments,the Dice scores of WT,TC,and ET regions are0.93705,0.90617 and 0.87470,respectively,indicating that the 2P-Res UNet-OMT model based on the VOMT algorithm proposed in this paper has a high accuracy for the segmentation results of brain tumors,and can accurately segment di(?)erent types of brains with di(?)erent shapes and sizes.The 2P-Res UNet-OMT model proposed in this work can substantially aid doctors in diagnosis and treatment,as well as present a novel research idea in medical imaging.