Detection Of Cerebral Microbleeds Based On 3D Neural Network

Cerebral microbleeds(CMBs)are important imaging and diagnostic biomarkers for cerebrovascular diseases and cognitive dysfunctions.Reliable detection of the location and amount of CMBs in brain tissue is crucial for the diagnosis,prevention and treatment of related diseases.In current clinical routine,CMBs are manually labeled by doctors from medical images,but this procedure is time-consuming and laborious,and the results are poor in repeatability.At present,there are some CMBs automatic detection methods developed by various image processing technologies.At present,there are some automatic detection methods developed by various image processing technologies.Most of the early methods mainly employed hand-crafted features based on shape and intensity information,and combined with traditional machine learning for detection.Due to the small size and complex features of CMBs,the detection effect of traditional methods is not ideal.Later,some scholars proposed to use deep learning for CMBs detection,and the effect has been greatly improved compared with the previous traditional detection methods.However,due to the insufficient use of the third dimensional spatial information with the traditional shallow neural network structure,the ability to extract three-dimensional features is limited,and there are still limitations in the detection accuracy and time performance.Three-dimensional neural network is a better solution,which is more consistent with the features of images and it can obtain the spatial information of the third dimension.On the basis of fully studying the principle and application of neural network,a two-stage cascade detection framework based on 3D neural network is adopted to realize candidate screening and accurate recognition of CMBs in magnetic resonance image.Firstly,in the first stage,a 3D full convolutional neural network was built to realize rapid and efficient large-scale screening,excluding a large number of background regions and obvious non-CMBs regions,and select candidates with high probability.Then in the second stage,3D AlexNet was used to extract the complex spatial features of CMBs,and the real CMBs and false positive samples were distinguished from the candidate regions selected in the first stage.At the same time,the network model is optimized.The batch normalization layer is added to the model structure to realize the normalization operation of the input data distribution in each layer of the network,thus speeding up the convergence of the training process and generalizing the network model.It can effectively alleviate the common "gradient dispersion" problem in the deep network,and improve the accuracy and performance of the model.In addition,the detection accuracy of input blocks with different sizes is compared to find out the input block size that can obtain the best detection effect.Finally,by comparing the experimental results with machine learning classification method and traditional convolution neural network method,it is proved that the proposed method can realize the automatic detection of CMBs in magnetic resonance images,effectively improve the detection sensitivity and accuracy,and reduce the number of false positive samples.