Research On Microvascular Decompression Image Segmentation Based On Deep Learning

Microvascular decompression in neurosurgical microsurgery has become the first choice for the treatment of trigeminal neuralgia and hemifacial spasm because of its small surgical trauma,few complications,high safety and good clinical effect.During the treatment of trigeminal neuralgia,hemifacial spasm and intracranial tumors,CTA,DSA and MRA are usually used to obtain cerebrovascular images.These images generally have limitations such as unclear small vessel imaging and lack of edge information.It is simpler and more convenient to obtain true color neurosurgical microsurgery images by endoscopy,and it has more important application prospects in clinical adjuvant therapy.At present,deep learning has been widely used in the medical field to assist doctors in completing surgical planning and clinical diagnosis.For neurosurgical microsurgery images,problems such as over-similarity in cerebrovascular classes,low contrast between cerebrovascular and brain tissue,and entanglement between cranial nerves lead to difficulties in cerebrovascular segmentation.Similarly,the segmentation of cranial nerves also faces such problems.In order to deal with the above challenges,this dissertation takes the deep learning method as the core,solves the difficulties of cerebrovascular and cranial nerve segmentation as the guidance,and designs the automatic segmentation algorithm as the goal to carry out in-depth research on the segmentation of neurosurgical microsurgery images.The main research contents and innovations are as follows :(1)In order to better integrate multi-scale features and make full use of global and local information,a semantic segmentation model of microvascular decompression image based on feature distillation is proposed.The model adds the Feature Distillation Block(FDB)to the encoder structure and the Atrous Spatial Pyramid Pooling(ASPP)module to the decoder structure to improve the semantic segmentation performance of microvascular decompression.FDB is added to the backbone network to further refine the extracted features.At the same time,Shallow Residual Block(SRB)is the main component of FDB,which enables the network to maximize the advantages of residual learning while ensuring lightweight.ASPP module is introduced into the decoder structure,and multi-scale features are fused to enhance the retention of features and boundary information,so that the boundary information of the target is more complete.The results show that the model can obtain more feature information and clearer target boundary,improves the semantic segmentation accuracy of microvascular decompression image,and provides help for future intelligent medical treatment.(2)Aiming at the problem of large network parameters and low semantic segmentation accuracy of real-time semantic segmentation of true color microvascular decompression images.A lightweight fast semantic segmentation network U-MVDNet for microvascular decompression scenarios is proposed,which is composed of an encoding and decoding structure.The Light Asymmetric Bottleneck Module(LABM)is designed in the encoder to encode the context features,and the Feature Fusion Module(FFM)is introduced in the decoder to effectively combine high-level semantic features and low-level spatial details.The experimental results show that for the microvascular decompression image test set,the parameter quantity of U-MVDNet on a single NVIDIA GTX 2080 Ti is only 0.66 M,the mean intersection over union reaches76.29%,and the speed reaches 140 FPS.And when the input image size is 640?480,U-MVDNet achieves real-time(24FPS)semantic segmentation on the embedded platform NVIDIA Jetson AGX Xavier.The model does not use any pre-training model,has a small number of parameters and fast reasoning speed,and its semantic segmentation performance is better than other comparison methods.It achieves a good trade-off between segmentation accuracy and speed.At the same time,it can also be easily developed and applied on embedded platforms with superior performance and easy deployment.(3)Aiming at the problem of low accuracy of some cranial nerves and cerebrovascular segmentation in current multi-class segmentation,a microvascular decompression image segmentation network based on cascade fusion(MCFNet)is proposed.The model uses Mobile Net as the backbone network,which has efficient inference speed and high-precision segmentation performance.At the same time,three modules are introduced: Cascade Fusion Module(CFM),Spatial Channel Attention Module(SCAM)and Feature Aggregation Module(FAM).The core module CFM combines different levels of features through feature pyramid module and feature cascade,and effectively aggregates the semantic and location information of cranial nerves or cerebrovascular.SCAM realizes the capture of key spatial information.The FAM effectively fuses the cross-level features,and the high-level semantic location information features are fused with the features from CFM to improve the segmentation performance of the model.The state-of-the-art segmentation methods are extensively evaluated and compared using microvascular decompression image datasets,ISIC-2016&PH2 and ISIC-2018 challenge datasets.Experiments show that MCFNet always outperforms the advanced methods and achieves the excellent segmentation performance.