sye@snu.ac.kr

To investigate the metal artifact reduction (MAR) performance of deep learning (DL)-based MAR technique in the evaluation of postoperative CT of total knee arthroplasty (TKA) patients

The training dataset consisted of 640 image pairs obtained from 10 lower extremity CT scans without a metal prosthesis. Each image pair consists of a metal artifact-free image with a virtual metal shape embedded in the original image and a metal artifact image simulated through sinogram handling. Our DL network is a convolutional neural network (CNN) with encoder-decoder structure and skip connections. The summation of MSE and SSIM losses were implemented for parameter updating. For the test dataset, we used 10 lower extremity CT examinations from 10 patients who had a previous history of TKA (7 patients with unilateral TKA; 3 patients with bilateral TKA), and a total of 13 knee joints were used for analysis. To evaluate the metal artifacts quantitatively, the area, mean attenuation, and artifact index (AI) within the dark streak artifacts were calculated in the original, O-MAR, and DL-MAR images. For qualitative analysis, images were rated with a 5-point Likert scale regarding the degree of overall metal artifacts, conspicuity of bone cortex and trabeculae, and assessment of soft tissue around the prosthesis. Continuous variables were compared between different MAR protocols using the repeated measures ANOVA and qualitative grading results were analyzed by using the Friedman test.

The O-MAR showed a 24% reduction in metal artifact area, while the DL-MAR showed an area reduction of more than 99%, almost completely eliminating the dark streak artifact. In terms of mean attenuation and AI, DL-MAR also showed better performance than O-MAR (P < 0.001). In qualitative analysis, DL-MAR showed significantly lower overall metal artifacts (P = 0.008) and better bone delineation (P = 0.020) compared to O-MAR. However, there was no significant difference in the assessment of soft tissue between two MAR protocols (P = 0.054), and DL-MAR showed unusual blurring of periarticular soft tissue.

The DL-MAR technique has been successfully developed and shown comparable performance with conventional projection completion algorithm.

The DL-MAR can effectively reduce severe metal artifacts caused by large TKA components, hence enabling its use in the diagnosis of postoperative complications of TKA.