Currently in clinical practice, tumor segmentation contributes to diagnosis and determination of radiotherapy area, leading to higher efficiency. In order to offer doctors help in lesion analysis and measurement, segmenting tumors from medical images is investigated in this paper. We propose a geometry information enhanced Unet to solve the problem of edge perception in medical images. In this network, we devise a method of extracting differential geometry information at the input of Unet, which makes full use of the tissue edge information of images to improve the segmentation accuracy of the network. This network makes the edge of tumors clearer by using Jacobian determinant and Laplace operator. Experiments on BraTS2018 dataset are performed to demonstrate that our network has superior performance to the baseline. And we apply our method to the clinical liver tumor CT data to explore practicality of our model in other tumor types or other medical image modality.

medical image, differential geometry, Unet, tumor segmentation