Comparison of UNet and DC-UNet models for an efficient segmentation and visualization of rodent hepatic vascular network from X-ray phase contrast imaging

This study proposes deep neural methods and tools for the extraction and visualization of vascular systems, through SR-PCT images (synchrotron radiation X-ray phase-contrast tomography) of murine liver. This is the first time that two deep learning architectures with different parametrizations were applied and compared for vessel segmentation with this imaging modality. Moreover, we propose to apply pre-processing steps (CLAHE, sigmoid and Gaussian filtering) in order to improve the contrast of raw data. We show that the best performance is obtained thanks to a DC-UNet model, learnt with these improved images. With this complete pipeline, we were able to segment and visualize in 3D the complete liver vasculature within a volume of more than 1000
³
voxels.