A Deep Learning Method For Automatically Delineating The Target Area Of Radiotherapy For Brain Tumors

Tumors seriously endanger the health of residents,and brain gliomas are a common malignant brain tumor that threatens the lives of patients.As the main auxiliary diagnosis and treatment method for brain tumors,Magnetic Resonance Imaging(MRI)provides important help for the diagnosis and treatment of brain tumors due to its rich diagnostic information.In tumor radiotherapy plan,accurate radiotherapy plan is the premise to ensure accurate treatment of patients,and target area delineation plays a key role in the radiotherapy plan.The delineation of the brain tumor target area and different brain tumor subareas from multi-mode MRI images can provide important information for the formulation of radiotherapy plans.The traditional method of manual segmentation is time-consuming and depends on the experience of the physician.The Convolutional neural network(CNN)of Deep learning performs well,which brings more ideas for the automatic and accurate delineation of brain tumors.The work in this paper is the pre-work of tumor radiotherapy system-tumor target area is automatically delineated,which provides the basis and guarantee for the accurate radiation therapy dose calculation and optimization,so as to be able to better implement a high-quality radiotherapy plan.The automatic target area delineation through the establishment of deep learning model can better help the implementation of radiotherapy plan.Different from natural images,MRI brain tumors are characterized by diverse and changeable shapes,complex structures,unstable positions,uneven gray scales of brain tissues,and great differences among different patients.Therefore,the automatic segmentation of brain tumors is a challenge,so the automatic delineation of brain tumor target area is of great research value.This paper proposes a deep learning technique based on multimodal MRI images to automatically and accurately segment brain tumors.The purpose of the study is to establish a deep learning segmentation model to automatically identify and segment tumor target areas from MRI multimodal images.The thesis first proposes a deep convolutional neural network segmentation model of brain tumors based on image blocks,and compares the experimental results of image blocks of different sizes,and then selects the image blocks of the best size for the experiment,so as to realize the automatic brain tumor segmentation.However,the model based on image blocks also has deficiencies,so the paper uses transfer learning as the theoretical basis and applies U-net network based on fully convolutional neural network to MRI brain tumor segmentation.The main content of the paper is as follows:(1)In view of the excellent performance of brain tumors and CNN,an MRI image brain tumor automatic segmentation model 6-CNN network based on image blocks was constructed,and the network was experimentally verified in the public data set BRATS 2015 to ensure the effectiveness of the network.Experimental results show that the network can adapt to the contours of brain tumors and achieve good segmentation results,indicating that the network can effectively segment brain tumors,but it also has the problems of complex training process,low segmentation efficiency,and easy loss of data space.(2)Due to the problems such as large storage space,complicated process and easy loss of spatial structure in 6-CNN network,and based on the theoretical basis of transfer learning,U-net network based on full convolutional neural network was introduced for segmentation of MRI image brain tumors.Different from the training method of image block adopted by 6-CNN network,U-net network adopts the whole image training to realize end-to-end training.The experimental results show that this method can not only effectively segment brain tumors,but also significantly improve the segmentation accuracy of each sub region within the tumor,and the segmentation efficiency is also significantly improved.