Research On Medical Image Segmentation Algorithms Based On Neural Network

Medical image segmentation is a key technology for analysis,understanding,and accurate quantitative diagnosis of medical images.It provides important auxiliary information for intelligent medical services such as clinical disease diagnosis,preoperative planning,intraoperative navigation and postoperative evaluation.Furthermore,its accuracy will directly affect the effect of diagnosis,treatment and postoperative evaluation.Although substantial achievements have been made in the research of semi-automatic segmentation with human-computer interaction and automatic segmentation,the diversity of modalities and complexity of medical images present many difficulties and challenges.In this paper,we focus on the problem of medical image segmentation in multiple complex scenes,and conduct an in-depth study of the medical image segmentation algorithm based on neural network.The specific content is summarized as follows:1.Brain Tumor Segmentation from MRI using PCNN based on Multifeature Grey Wolf OptimizerThe segmentation performance of traditional methods such as thresholding and clustering is susceptible to the uneven image grayscale and noise,and the accuracy is low in MRI brain tumour segmentation with blurred boundary,low contrast and high noise.To solve this problem,this paper proposes an improved pulse coupled neural network(PCNN)based on the grey wolf optimizer with multiple hybrid features for medical image segmentation.Since a single modality can only reflect limited tissue information,which cannot effectively extract rich features from brain tumor regions,a two-stage framework is proposed including image fusion and image segmentation.In the stage of image fusion,rich information from different modalities is fused by utilizing a multimodal medical image fusion model based on the maximum energy region.In the stage of image segmentation,based on the meta-heuristic swarm intelligence search algorithm,an improved multiple hybrid features grey wolf optimizer-based PCNN segmentation model is designed,which can adaptively set the parameters of PCNN according to the features of the image.Experimental results prove that,the proposed algorithm achieves higher segmentation accuracy in MRI brain tumor segmentation.2.Multi-scale Context-aware Network for Medical Image SegmentationIn response to the complex features of medical images in different segmentation tasks(e.g.,modal diversity,class imbalance,the variable shape and size of organs or lesion regions),and the limitations of segmentation methods based on deep convolutional neural networks in context extraction and detailed information preservation,this paper constructs a multi-scale context-aware network(CA-Net).The multi-scale context fusion module(MCF)is introduced in the last layer of the encoder,which can capture the context information in deep semantic feature maps from multiple scales of spatial and channel.Meanwhile,more spatial granularity information from feature maps of each shallow encoder is preserved by introducing dense skip connections between the encoder and decoder.Experimental results demonstrate that CA-Net achieves the best segmentation performance without any pre-processing and post-processing operations in complex medical image segmentation tasks with different modalities,and also effectively reduces the number of model parameters and FLOPs.3.Multi-scale Context-aware and Semantic Adaptor Network for Medical Image SegmentationIn order to further enhance the feature representation capability of the network in complex scenarios,such as blurred boundaries of lesion regions,low contrast with surrounding tissues,uneven brightness,and interference from similar background colors,a deep encoder-decoder network based on multi-scale context-aware and semantic adaptor(MCSA-Net)is proposed.It can extract rich semantic features from deeper layers of the images,and meanwhile reduce the adverse effects on direct fusion of semantic features from deep and shallow layers in the skip connections.On the one hand,the multi-scale context-aware module(MCAM)is adopted in multiple deep layers of the encoder to enhance the representation ability of feature learning at different scales in deep semantic feature maps and suppress interference in non-target regions.On the other hand,the multi-level semantic adaptor module(MSAM)is utilized to capture more spatial details from adjacent encoder feature maps,and refine the boundaries of the target area,as well as eliminate the differences of abstract semantic features between the encoder and decoder.Experimental results prove that the algorithm based on multi-scale adaptive learning and semantic adaptor not only improves the segmentation accuracy of the boundary,but also effectively suppresses the interference of the surrounding similar background information.4.Saliency Guidance and Uncertainty Supervision Encoder-Decoder Network for Medical Image SegmentationTo address the problems that under-segmentation or mis-segmentation still exists in the segmentation of some complex samples in previous works(Algorithm 2 and Algorithm 3),for example,large differences in internal features of tissues,organs or lesion regions while small differences in features between classes,complex textures,and more interference information in the background,this paper proposes a semantic segmentation algorithm for medical images based on saliency guidance and uncertainty supervision.The saliency map and uncertainty probability map are generated by the saliency-guided module(SGM)and the uncertainty-supervised module(USM)respectively.Then the two maps together with original image are fed into the segmentation network as supervised information,guiding the network to focus on learning features of target regions,moreover,enhancing the representational capacity of the uncertain classification regions and complex boundaries.Experiments show that the proposed algorithm can better locate the boundaries of target regions,and has better generalization and robustness by learning the features of the uncertain classification region,which to some extent solves the shortcomings of previous work in the above complex segmentation scenarioes.To sum up,in view of the current difficulties and challenges in medical image segmentation,using neural networks as an entry point,this paper presents an in-depth study on feature analysis of medical images,segmentation algorithm design and model construction,effectively improving the segmentation performance of medical image segmentation tasks in multiple complex scenes.