Mitigating metal artifacts in cone-beam CT images through deep learning techniques
10.3760/cma.j.cn112144-20231030-00233
- VernacularTitle:基于深度学习的锥形束CT影像金属伪影去除研究
- Author:
Linghui JIA
1
;
Honglei LIN
;
Songwei ZHENG
;
Xiujiao LIN
;
Dong ZHANG
;
Hao YU
Author Information
1. 福建医科大学附属口腔医院修复科，福州350002
- Keywords:
Cone-beam computed tomography;
Convolutional neural network;
U-net;
Metal artifacts;
Metal artifact reduction
- From:
Chinese Journal of Stomatology
2024;59(1):71-79
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop and evaluate metal artifact removal systems (MARSs) based on deep learning to assess their effectiveness in removing artifacts caused by different thicknesses of metals in cone-beam CT (CBCT) images.Methods:A full-mouth standard model (60 mm×75 mm×110 mm) was three-dimensional (3D) printed using photosensitive resin. The model included a removable and replaceable target tooth position where cobalt-chromium alloy crowns with varying thicknesses were inserted to generate matched CBCT images. The artifacts resulting from cobalt-chromium alloys with different thicknesses were evaluated using the structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR). CNN-MARS and U-net-MARS were developed using a convolutional neural network and U-net architecture, respectively. The effectiveness of both MARSs were assessed through visualization and by measuring SSIM and PSNR values. The SSIM and PSNR values were statistically analyzed using one-way analysis of variance (α=0.05).Results:Significant differences were observed in the range of artifacts produced by different thicknesses of cobalt-chromium alloys (all P<0.05), with 1 mm resulting in the least artifacts. The SSIM values for specimens with thicknesses of 1.0 mm, 1.5 mm, and 2.0 mm were 0.916±0.019, 0.873±0.010, and 0.833±0.010, respectively ( F=447.89, P<0.001). The corresponding PSNR values were 20.834±1.176, 17.002±0.427, and 14.673±0.429, respectively ( F=796.51, P<0.001). After applying CNN-MARS and U-net-MARS to artifact removal, the SSIM and PSNR values significantly increased for images with the same thickness of metal (both P<0.05). When using the CNN-MARS for artifact removal, the SSIM values for 1.0, 1.5 and 2.0 mm were 0.938±0.023, 0.930±0.029, and 0.928±0.020 ( F=2.22, P=0.112), while the PSNR values were 30.938±1.495, 30.578±2.154 and 30.553±2.355 ( F=0.54, P=0.585). When using the U-net-MARS for artifact removal, the SSIM values for 1.0, 1.5 and 2.0 mm were 0.930±0.024, 0.932±0.017 and 0.930±0.012 ( F=0.24, P=0.788), and the PSNR values were 30.291±0.934, 30.351±1.002 and 30.271±1.143 ( F=0.07, P=0.929). No significant differences were found in SSIM and PSNR values after artifact removal using CNN-MARS and U-net-MARS for different thicknesses of cobalt-chromium alloys (all P>0.05). Visualization demonstrated a high degree of similarity between the images before and after artifact removal using both MARSs. However, CNN-MARS displayed clearer metal edges and preserved more tissue details when compared with U-net-MARS. Conclusions:Both the CNN-MARS and U-net-MARS models developed in this study effectively remove the metal artifacts and enhance the image quality. CNN-MARS exhibited an advantage in restoring tissue structure information around the artifacts compared to U-net-MARS.
