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) image is relevant for several clinical applications. However, the liver segmentation on noisy images obtain incorrect liver vessel segmentation which may lead to distortion in the simulation of cooling effect near the vessels during the planning. In this study, we present a framework that consists of three well-known and state-of-the-art denoising techniques, Vesssel enhancing diffusion (VED), RED-CNN, and MAP-NN and using a state-of-the-art Convolution Neural Networks (nn-Unet) to segment the liver vessels from the CECT images. The impact of denoising methods on the vessel segmentation are ablated using with multi-level simulated low-dose CECT of the liver. The experiment is carried on CECT images of the liver from two public and one private datasets. We evaluate the performance of the framework using Dice score and sensitivity criteria. Furthermore, we investigate the efficient of denoising on roughness of the surface of liver vessel segmentation. The results from our experiment suggest that denoising methods can improve the liver vessel segmentation quality in the CECT image with high low-dose noise while they degrade the liver vessel segmentation accuracy for low-noise-level CECT images.