Research On Multi-label Image Segmentation Model Based On Mouse Brain

Medical imaging has developed rapidly in recent years,medical images are an important means of clinical diagnosis,fully mining the information contained in medical images is of great significance.The combination of deep learning and medical images has solved many problems that the medical world cannot break through.The 3D digital reconstruction of the mouse brain from the 2D image stack is known for its complexity and time-consuming nature,which is due to the extremely high packing density of the vasculature and neural network in the brain.The three-dimensional reconstruction of the local dense mapping of the neuron network in the cerebral cortex is the key to understanding the structure,function and development of the brain circuit.The MOST data set is investigatedby depth model,in order to better assist the 3D reconstruction of the mouse brain in the biological community.In this thesis,we use deep learning to combine with our data set,and we propose a Unet-based RI-Unet segmentation model for biomedical images to perform dense volume segmentation.The RI-Unet and many classic models are compared in the MOST data set.The RI-Unet framework is used in the MOST dataset to achieve efficient and accurate segmentation of somatic cells and blood vessel structures in the mouse brain.Due to the inherent category difference in MOST data,a weighted loss function is proposed to improve the model’s segmentation accuracy on the MOST data set and to solve the problem of category imbalance.We use cross-detection technology to eliminate some outliers in the data set and minimize the impact of data set errors.A large number of experiments have been done on the MOST data set.The experimental results show that the model,loss function design and outlier detection design proposed in the article greatly improve the accuracy of segmenting the mouse brain,and also increase the training speed.The performance of RI-Unet is 0.9948(recall rate),0.9953(accuracy)and Dice coefficient 0.9845(Somata),0.9510(Vessel),among which Dice coefficient increases by 1% and 4% respectively.The segmentation result is greatly improved.The results show that RI-Unet is superior to the existing deep convolution model for segmenting mouse brain vessels,and we get a more generalized model.The prediction results of the larger MOST image stack can be merged into a block organization,which is used for the 3D digital reconstruction of the brain of the mouse,which is of great significance for the study of neurovascular networks.