Automatic segmentation model combining U-Net and level set method for medical images

Abstract

We introduce a new method that combines a constrained term and level set method for the automated segmentation of medical image. There are two types of constrained terms, fully automatic and semi-automatic. It is fully automatic to use the U-Net’s segmentation result as a constrained term, and the manual segmentation result as a constrained term is semi-automatic. The level set method does not require a large training set and is theoretically very explanatory, but is usually sensitive to the initial contour. The U-Net can segment more complex medical images, but requires a large number of manually labeled images and usually needs to be normalized to produce a good generalization. Therefore, the combination of these methods combines the advantages of both methods, resulting in a method that requires a small training set and produces accurate segmentation results. We test our method on the melanoma and left ventricle images. Among them, when segmenting melanoma images, our semi-automatic segmentation and full-automatic segmentation results are better than the U-Net and RSF segmentation results alone. When segmenting the left ventricle images, our semi-automatic segmentation result is better than the RSF segmentation result.