Research On Deep Learning Based Medical Image Enhancement And Segmentation Algorithm

In recent years,with the continuous improvement of the national medical level and the continuous popularization of medical imaging equipment in grassroots hospitals,medical image data has become an important basis for doctors to make pathological diagnosis.The processing of medical image data by computer technology has also aroused great interest of researchers.With the continuous development of deep learning technology and the continuous improvement of computer hardware,medical image assisted diagnosis technology based on computer technology has made great progress.Due to the influence of machinery and environment in the process of medical image acquisition,noise and metal artifacts are often introduced into the image to damage the quality of the image.Therefore,medical image enhancement is an important pre-processing work for further medical image analysis.Automatic segmentation of regions of interest（ROI）in medical images using relevant algorithms has important research significance and clinical application value for improving the efficiency of professional doctors in reading images,assisting doctors in clinical diagnosis and relieving the intensity of manual segmentation.In this paper,the enhancement and segmentation algorithms of MRI images are studied.This paper mainly studies the MRI image denoising algorithm based on the encoder-decoder network,the dual-supervised MRI denoising network based on the attention mechanism and the MRI brain tumor segmentation network based on the multi-scale strategy and the attention mechanism.The main work and contributions are as follows:（1）Propose an encoder-decoder based deep neural network for MR image denoisingIn this algorithm,based on the depth of the neural network can be regarded as a kind of implicit regularization function point of view,and in the new design of the codec based network as the premise of implicit regularization,introduce the concept of image structure similarity,build a new data fidelity term,and the image resolution is used to calculate a priori and designed a kind of regularization item（penalty）.The new objective function of model optimization is composed of the new data fidelity item and the regularization term based on clarity calculation.The proposed deep neural network consists of three main parts:encoder network,decoder network and jump connection structure.The encoder network consists of five down-sampling modules,which extract low-resolution or relatively abstract image features.The structure of the decoder network is similar to that of the encoder network,which is composed of five up-sampling modules in series to perform prediction and recovery of higher resolution image features.The function of jump connection structure is to transfer abstract information directly from the encoder to the decoder,which enhances the information recovery ability of the decoder and facilitates the generation of finer image features.（2）Propose dual Supervisions attention based encoder-decoder network for magnetic resonance image denoisingBased on the traditional encoder-decoder network,a parallel decoder network denoising method based on ₁l and SSIM loss supervision is proposed.Among them,the loss function adopted can avoid speckle artifacts,and effectively enhance the ability of the network to capture texture information,and guide the decoder sub-network to generate images from different angles.In order to integrate the outputs of different decoders,an attention mechanism model is introduced.The attention model uses the hidden feature maps extracted from each stage of the encoder,and learns to generate the weight map after splicing and fusion of the feature images.The weight map is used to weight the output of the two decoders.Finally,the convergence of the whole network is maintained by using the mean square error loss.（3）Propose a multi-scale strategy based 3D dual-encoder brain tumor segmentation network with attention mechanismWe propose a medical image segmentation model,introduce multi-scale strategy and attention mechanism,and design a multi-scale 3D NMR image segmentation network with dual encoders.This network firstly adjusted the input 3D MRI images to two scales,and then used the dual encoder network to extract the features and capture the context semantic information of the 3D MRI images at the two scales.In each encoder network,the features of different locations in the network are extracted,which come from multiple locations in the network and contain rich context information.Then,the features are spliced and fused as the input feature graph of the proposed attention model.Then design the attention model to study the scale characteristics of the weight of corresponding graph,small scale to large objects and large scales of small objects give big weight respectively,and the different scales of weight features a pixel count weight figure,and the weights of mapping is applied to the output of the encoder network,multi-scale encoder to generate network characteristics combined.Then,a decoder network is designed to accurately locate or mask the prediction with the sum of features.In order to enable the decoder network to assemble more accurate predictions,jump connections are used between a large encoder network and a decoder network to transmit high-resolution features.