HyperUNet for Medical Hyperspectral Image Segmentation on a Choledochal Database

Abstract

Microscopy medical hyperspectral images, which are characterized in multiple observation bands under different spectral frequencies, contain profuse spectral information for disease diagnosis. Consequently, an increasing number of deep learning methods have recently been proposed to solve the medical hyperspectral image segmentation task. In this study, we propose a new segmentation network (HyperUNet) as a better version of UNet for medical hyperspectral image segmentation on a choledochal database. Considering the useless spectral in-formation that exists in the hyperspectral image that is irrelevant to our task, HyperUNet first uses the linear transformation block to extract the useful spectral information from the hyperspectral image, and then applies the UNet model to it to capture the tumor area. Finally, when reconstructing the mask, HyperUNet applies the multi -scale loss function in cases of underuse and overuse of low-level detailed features and high-level semantic features. We compare our HyperUNet to other competing methods, and the results show that our HyperUNet is superior.