Infant Brain MR Image Segmentation Based On Modified Fully Convolution Network

Neuropathic diseases are extremely difficult to cure,which threatens the health and safety of hundreds of people every year.According to the research,autism is related to the development of the brain during the birth period,so the early brain imaging analysis can provide valuable reference information for the early data.Nuclear magnetic resonance imaging(MRI)is widely used in clinical diagnosis because of its non-invasive,non-invasive and high contrast characteristics.However,due to the low contrast between brain tissues in baby MR images during growth and development,traditional segmentation methods are difficult to obtain more accurate results.Deep learning is widely used in image segmentation because of high robustness and analysis efficiency.However,because the analysis of medical images requires high accuracy,especially baby MR image,the traditional deep learning method is difficult to meet the practical application.In view of the low accuracy of traditional deep learning methods,two fully convolutional network models are proposed in this paper.The specific research contents are as follows:(1)Aiming at the excessive loss of details in Image segmentation by U-Net,a deep learning model combining residual connection and dense connection with U-Net model is proposed.The model is composed of two blocks,the residual block can alleviate gradient disappearance and optimize training,Dense block can enhance feature propagation,and U-Net framework can realize end-to-end segmentation of pixel.The three complement each other to improve segmentation accuracy.(2)According to the characteristics of MR images of multimodal infant brain,a deep learning model combining multiple inputs,residual connection and multi-scale convolution is proposed on the basis of U-Net.In the model,the multi-input structure can extract features from multiple directions;The residual multi-scale block can alleviate the gradient disappearance and extract richer features;The fusion block enhances the use of shallow features.In the subsequent experiments,the 2.5D input further improved the performance of the model,and the segmentation performance of adult human brain MR images verified the applicability of the model in monomodal data.Finally,the contrast experiment of the image segmentation correlation loss function demonstrate that cross-entropy is the model’s best loss function.By comparing the segmentation results of clinical infant brain MR images,the two methods proposed in this paper have achieved more accurate results in the segmentation of elongated structures and irregular boundaries.This paper analyzes and verifies the effectiveness of the two methods from the perspectives of visualization and quantification.