Research On Brain MRI Segmentation And Classification Based On Deep Learning And Vector Quantization

Accurate and fast segmentation and classification of medical images is a critical step in medical diagnosis.Effective segmentation and classification results can reduce the burden on doctors and improve the accuracy of diagnosis,which can help to finalize the treatment plan.As the brain is the most important organ in the human body,segmentation and classification of brain MR images are especially important.In recent years,supervised image segmentation methods based on deep learning have made great progress and achieved excellent performance in various fields of segmentation tasks.However,when the data processed by such methods differ significantly from the distribution of the training data,the performance can be significantly degraded.Some unsupervised segmentation methods,however,show better robustness when the distribution of the data differs significantly from the training data,but the performance of such methods is still limited.Therefore,how to achieve accurate segmentation and classification of brain MRI is the focus of this thesis.In this thesis,we propose a brain MRI segmentation and classification method based on deep learning and vector quantization to accomplish automatic segmentation and classification tasks and achieve high performance.First,to solve the problem of small input image size under the limitation of GPU memory,this thesis designs an asymmetric 3D U-Net,which enables the model to accept full-size images as input to learn more global image semantic information.When applied to new data,the model can provide more accurate segmentation and classification results and more accurate information for the subsequent steps.Secondly,to obtain voxel point information in the edge vector and accurate edge information,this thesis proposes a 3D Sobel-based edge extraction and pattern division scheme to achieve edge enhancement and extraction.When sub-blocks patterns are divided,the edge voxel point location information is recorded and provided to the edge sub-blocks for quantization;finally,to enhance the feature representation of non-edge vectors,this thesis performs a deep belief network by a self-supervised approach and uses a trained encoder for compressed feature representation of the input samples.This method leads to a further improvement of the final segmentation performance.This thesis conducts experiments on three 3D brain MR image datasets,and the experimental results are analyzed in detail.The experimental results show that the method in this paper can segment and classify the input data accurately.The proposed method achieves the best performance among unsupervised methods and has competitive performance when compared with supervised methods,which is more practical.Finally,the experiments fully validate the effectiveness of the proposed key techniques.