Brain Image Segmentation Based On Deep Learning

Brain is the most important and complex organ in human body.Brain diseases will pose a major threat to human health.Medical imaging technology can acquire brain tissue images in a non-invasive way.Segmentation of different tissues and organs in brain medical images is of great significance to improve the diagnostic ability of medical diagnostic personnel for normal and diseased organs.Artificial segmentation of different tissues and organs in the image is time-consuming and laborious,and requires a high level of diagnosis.By collecting a large number of medical images,this paper designs a fast location algorithm and segmentation algorithm for brain medical images,and realizes an automatic segmentation system for brain medical images.The main work of this paper is as follows:1.Construct medical image data sets,he data set consists of two parts:552 human computed tomography image sequences?CT?and 140 tagged magnetic resonance imaging sequences?MRI?with four modes:T1,T1₁mm,T1_IR and T2_FLAIR.2.A rapid positioning algorithm for brain image based on target detection is designed and implemented,and a multi-scale detection network based on Yolo is constructed.The sagittal section of medical image is processed by dense projection,maximum connected area and threshold segmentation.After training by cutting brain image center and rotating data enhancement,the model is used for CT image.The average IOU of image classification reached 96.4%,and the average prediction time of single amplitude remained at 0.25s.Among them,the average IOU of image classification for brain region was 99.1%.3.A brain image segmentation algorithm based on improved U-net network is designed and implemented.The network architecture is adjusted according to the specificity of medical image based on U-net network.A cascaded network composed of 19 convolution layers,16 activation layers and 4 deconvolution layers is designed.The centroid rotation and noise-added data enhancement method are combined with location and classification loss function training.After that,the average test accuracy of the model for brain MRI image segmentation is 92.7%,and the average prediction time is 3.10 seconds.4.A brain image segmentation algorithm based on self-built neural network model is designed and implemented.A multi-input residual network consisting of 26 convolution layers,24 normalization layers,3 up-sampling layers and 3 Maximum pooling layers was designed for the specificity of medical images.After training with the method of joint input of multi-modal MRI data and class-based random data enhancement,the average test accuracy of brain MRI image segmentation reached 98.2%,and the average prediction time was 3.90s.