Soma Segmentation Using U-Shaped Convolutional Neural Networks with Weight Boundary Mechanism

Abstract

Soma segmentation is crucial for subsequent analysis of brain images, and can be used for analysis of neuronal morphology and other aspects. Although the results of manual soma segmentation are more accurate, this method is very time-consuming and labor-intensive, so an accurate method of automatic soma segmentation is needed. Existing optical imaging images of the brain tend to be large, and one way we segment somas is to take the approximate coordinates of the center of the soma we need, then cut the image into small cubes one by one, and segment the somas in the small cubes. However, in this cube, in addition to the soma roughly located in the center that we want to segment, there are often other somas in the periphery, especially when they are close to our target soma, and some of them are only left incomplete during the cutting process. In addition, due to some defects of optical imaging and other problems, it leads to many images with low signal-to-noise ratio, which severely affects the segmentation effect. Therefore, we incorporate two mechanisms in the training of U-shaped convolutional neural network to alleviate the above problems, one is soma edge identification, and the other is the enhancement of multi-soma construction with soma shifted. Compared with existing popular medical image segmentation models, our method is significantly better in soma segmentation.