MSD-Net: A Multi-Scale Semantic Segmentation Method for Images with Metal Artifact Interference

The semantic segmentation for computed tomography (CT) images is an important step in clinical diagnosis. Metal artifacts may cause significant segmentation errors, and even after artifact reduction, conventional methods still struggle to achieve accurate segmentation results. To address the reduction in segmentation accuracy caused by artifact interference, this paper proposed a Multi-Scale Detail Enhanced Network (MSD-Net). The network employed a Channel Multi-Scale module and a Mamba module to extract multi-scale information and selectively capture input information, establishing clear relationships between global and local features. When processing blurred edge information, the proposed multi-scale detail enhancement structure enhanced the extracted multi-scale detail information, and used a multi-scale feature fusion method to extract semantic information at more appropriate scales. High-precision segmentation results were achieved on images following metal artifact reduction.

I. Introduction

Computed tomography (CT) has become an indispensable diagnostic tool in recent years[1]. Accurate organ segmentation in CT images provides important support for disease diagnosis and quantitative analysis, surgical planning and navigation, monitoring of treatment effects, and various other clinical tasks. [2]. Semantic segmentation methods for CT images without metal implants have already achieved outstanding results. However, when metal artifacts are present in CT images, these artifacts can significantly degrade image quality, leading to numerous errors in the segmentation results. This issue is particularly challenging when metal implants of varying shapes and sizes are present in the body. Even after artifact reduction, CT images still contain metal implants and incomplete tissue structure edges. Conventional segmentation methods struggle to recognize the relevant features of metal and incomplete tissue structure edges in the images, leading to the misclassification of metal as tissue and the inability to effectively identify incomplete edges. Therefore, developing a segmentation method for CT images with reduced metal artifacts is an important research direction.