Segmentation And Three-dimensional Reconstrucetion Of Brain Tumor MRI Images

The incidence and lethal rate of brain tumors have increased year by year,and sufferers range from young to old.The prevention,diagnosis and treatment of brain tumors become more and more important.Traditional methods rely on doctors to judge brain MRI slice images,which are easily affected by human factors and inefficient.Computer image processing and deep learning provide a new boost for the diagnosis and treatment of brain tumors.The trained convolutional neural network can quickly and steadily segment brain tumors MRI images,assist doctors in diagnosing brain tumors,and improve the accuracy and speed of segmentation.The three-dimensional reconstruction technology can provide the intuitive 3D visualization of brain tumors,which can help doctors discuss patient's condition and formulate appropriate therapeutic schedule.In this paper,two datasets,the clinical dataset and BraTS2017 dataset,are used as experimental data.Based on the different datasets,three tasks have been carried out:denoising,segmentation and 3D reconstruction of brain tumors.The main work of this paper is as follows:(1)Denoising of brain tumor images.Due to lack of manual preprocessing except labelling the tumor,there are non-zero black background noise and bias field noise in clinical dataset,which will affect the effect of subsequent segmentation.Therefore,a method of combining the Otsu method and morphological operation is used to remove the non-zero black background,and the N4ITK method is used to correct the bias field of the clinical dataset.The results show that the algorithms perform well.(2)Segmentation of brain tumors.This paper uses convolutional neural networks for segmentation.The typical CNN models of segmentation such as FCN,U-Net,SegNet,Attention U-Net,and FRRN are investigated and reproduced.Based on the inspiration of these models,an enhanced U-Net model is proposed.In the enhanced U-Net,residual connection is adopted to improve convolution layer in upsampling path;bilinear interpolation is used to substitute for deconvolution and unpooling in upsampling path;the "one-to-many" denser skip-connection,in which strided dilated convolution is used for downsampling,is designed to replace the "one-to-one" parallel skip-connection.The clinical dataset and the BraTS2017 dataset are both used to train and evaluate each model,and the effectiveness of enhanced U-Net is proved through comparative experiments.(3)Three-dimensional reconstruction of brain tumors.This paper studies the common methods of 3D reconstruction technology,and chooses to use Marching Cubes method in surface rendering and Ray Casting method in volume rendering to conduct 3D reconstruction of the BraTS2017 dataset.The reconstruction results show the practicability of methods.