Impact of Image Denoising Techniques on CNN-based Liver Vessel Segmentation using Synthesis Low-dose Contrast Enhanced CT Images

Abstract

Liver vessel segmentation in contrast-enhanced CT (CECT) images has a significant role in the planning stage for liver cancer treatment, such as radiofrequency ablation (RFA). Lowering the radiation dose in CECT imaging to reduce radiation risk to the patient degrades the quality of the image and potentially affects the liver vessel segmentation. In recent years, the convolutional neural network (CNN) has shown significant achievement in medical image analysis, including segmentation and denoising tasks. This paper presents a study on a new framework consisting of three well-known denoising techniques, including vessel enhancing diffusion (VED), RED-CNN, and MAP-NN, along with the state-of-theart segmentation method (nnU-Net) to segment the liver vessels in CECT images. We quantitatively evaluate the impacts of denoising techniques on the vessel segmentation on multi-level simulated low-dose CECT images of the liver. The performances of the liver vessel segmentation method combined with the denoising techniques are evaluated using Dice score, sensitivity metric, and processing time. In addition, the effect of denoising on the surface roughness of the segmented liver vessel is also investigated. The experiments show that the image denoising techniques improve the quality of liver vessel segmentation on high noisy CECT images while also reducing the segmentation accuracy on low-noise-level CECT images.