Research On Segmentation Method Of Multimodal Brain Tumor MR Image Based On 3D U-Net

When using a computer for auxiliary diagnosis,the segmentation of medical images is a basic and critical task.Due to the characteristics of glioma itself,the MR images of brain tumors show significant inconsistencies and irregular shapes,so it is still a challenging task to perform fine segmentation of brain tumors in medical images.Traditional image segmentation algorithms have been difficult to meet the relevant needs of current clinical applications.With the development of deep learning,convolutional neural networks have begun to be used in brain tumor segmentation tasks.For brain tumor segmentation,this thesis takes multimodal brain tumor MR images as the research object,and uses deep learning-based segmentation methods as technical means to carry out the following research:(1)The brain tumor MR images were processed to construct a multi-modal brain tumor MR dataset.Adaptive denoising and normalization are performed on the existing brain tumor MR images to improve the contrast between the tumor area and the background,preserve the details of the lesions,complete and reasonable MR image data of brain tumor were constructed for model training and validation.(2)3D U-Net with Trans-coder for Brain Tumor Segmentation.Aiming at the problem of over-segmentation or inaccurate segmentation of some tiny tumor regions and tumor boundaries by 3D U-Net network,a 3D U-Net brain tumor segmentation algorithm based on Trans-coder is proposed.The algorithm is based on the three-dimensional UNet network and uses the effectiveness of the attention mechanism to design a Transcoder module containing multi-head self-attention,which is embedded in the encoder end of the U-Net network to improve the performance of the segmentation effect while improving the performance of the segmentation effect.Reduce the amount of computation.At the same time,a variable autoencoder is used for regularization to prevent overfitting.Experiments show that on the multimodal brain tumor MR dataset,the method proposed in this thesis has better results in avoiding over-segmentation and improving segmentation details than other methods.(3)Considering the computational complexity and the three-dimensional features of brain tumor images,the thesis further introduces the axial attention mechanism into the three-dimensional U-Net,and uses the attention mechanism to extract more abundant tumor image features from the three dimensions of length,width,and height.The experimental results show that the improved segmentation network can retain multi-level features and extract richer multi-scale features for tumor image segmentation.Compared with other methods,the proposed method can segment more accurate and complete tumor lesions.