Metal artifact reduction combined with deep learning image reconstruction algorithm for CT image quality optimization: a phantom study

Phantom

In this study, the Lungman chest phantom (Lungman ph-1, Kyoto Kagaku Inc., Japan) was utilized. The anatomical structures, including the trachea, pulmonary vessels, and mediastinum, were simulated using tissue substitutes. Thirteen spherical nodules (CT value = −800 HU, corresponding diameters = 12, 10, 8, and 5 mm; CT value = −630 HU, corresponding diameters = 12, 10, 8, 5, and 3 mm; CT value = 100 HU, corresponding diameters = 12, 10, 8, 5 mm) were randomly placed in the phantom using cotton. To investigate the impact of metal artifacts, a pacemaker was attached to the upper left of the chest phantom (Fig. 1A). Four non-solid pulmonary nodules (PNs) were obscured by streak artifacts (Fig. 1B).

Image acquisition and reconstruction

All scans were conducted using a 256-row multidetector CT scanner (Revolution Apex CT, GE Healthcare). To investigate the impacts of various scanning conditions on metal artifact reduction, three tube voltages (70, 100, and 120 kVp), along with their corresponding tube currents, were selected to achieve standard (3 mSv) and low (0.5 mSv) effective doses. The remaining scanning parameters were fixed across all scanning scenarios, as follows: display field of view (DFOV) of 42 cm × 42 cm, a pitch of 0.992, a detector width of 80 mm, a rotation time of 0.8 s/r, and a slice thickness of 1.25 mm. Furthermore, all acquisitions were reconstructed using high strength DLIR (DLIR-H), DLIR-H with metal artifact reduction (DLIR-H MAR), 50% adaptive statistical iterative reconstruction-V (ASIR-V), and ASIR-V with metal artifact reduction (ASIR-V MAR). CT scanning was repeated three times for each scenario.

Objective image quality evaluation

For quantitative analysis, the artifact index (AI), background noise, signal-to-noise ratio (SNR) of artifact-impaired non-solid PNs, noise power spectrum (NPS) of artifact-free regions were calculated. All image sequences were loaded into MITK software (v2024.06, The German Cancer Research Center, Heidelberg, Germany), and region of interests (ROIs) were delineated in background air, artifact areas, and artifact-impaired PNs by a well-experienced radiologist (with nine years of chest radiology experience) based on 3 mSv 120 kVp DLIR-H MAR images. To ensure consistent quantitative analysis of the same ROIs in other images, these ROIs were saved and subsequently imported into other images for analysis. According to a previous study (Chae et al., 2020), the AI was quantified using the following formula: $\mathrm{AI}=\sqrt{{\mathrm{SD}}_{\text{artifact}}^2-{\mathrm{SD}}_{\text{background}}^2}$, where SD_artifact and SD_background represent the standard deviation (SD) of the streak artifact and background, respectively. To represent the extent of the artifacts as comprehensively as possible, ROIs (50 mm²) were placed in five consecutive artifact-pronounced slices. Background ROIs (50 mm²) were located in the air among five consecutive artifact-free slices, and the SD_background in the AI formula was the average of these five background SD. To assess the influence of artifacts on PNs, the SNR values of artifact-impaired non-solid PNs (Fig. 1B) were calculated using the following formula: $\mathrm{SNR}=\frac{\left|{\text{mean}}_{\mathrm{PN}}\right|}{{\mathrm{SD}}_{\mathrm{PN}}}$, where mean_PN and SD_PN refer to the average and SD of the CT values of the four PNs at the maximum slice, respectively. To investigate the influence of DLIR and MAR algorithms on the image texture of artifact-free regions, NPS was evaluated in a homogeneous heart using imQuest software (Clinical Imaging Physics Group, Duke University, Durham, NC, USA) and the NPS area, average spatial frequency (f_avg) and peak spatial frequency (f_peak) were calculated.

Subjective image quality evaluation

Two radiologists (5/9-year-experience in CT image diagnosis) independently performed subjective evaluations using a 5-point scale to assess the extent of metal artifacts, the confidence in diagnosing artifact-impaired PNs, and the overall image quality. The extent of metal artifacts was rated as follows: severe artifacts, unable to be diagnosed = 0, pronounced artifacts = 1, moderate artifacts = 2, mild artifacts = 3, and no artifacts = 4. The confidence in diagnosing artifact-impaired PNs was graded as follows: undetectable = 0, poorly detectable = 1, moderately detectable = 2, well detectable = 3, and manifestly detectable = 4. The overall image quality was rated as follows: very poor = 0, poor = 1, acceptable = 2, good = 3, and excellent = 4.

Statistical analysis

IBM SPSS statistical software (version 25.0, IBM Corp) was used for statistical analyses. According to the Shapiro–Wilk test, the objective parameters and subjective scores did not exhibit the normality. Therefore, AI, noise and SNR were expressed as M [Q1, Q3], where M represents the median, and Q1 and Q3 denote the first quartile and third quartile, respectively. The differences in these parameters among various image sets were compared using the Kruskal–Wallis test with a Bonferroni post hoc test. The evaluation of subjective consistency between two radiologists was conducted using Cohen’s kappa test, with values greater than 0.75 indicating high consistency, values ranging from 0.4 to 0.75 indicating average consistency, and values less than 0.4 indicating poor consistency. p < 0.05 indicates statistically significant differences.

Objective image quality evaluation

Table 1 summarized the results of noise and AI evaluations of various reconstruction algorithms and scanning scenarios (radiation doses and tube voltages). Compared to ASIR-V, the background noise of DLIR-H was reduced by 25.95% to 53.50% (paired calculation), whereas the background noise of DLIR-H MAR was reduced by 27.73% to 54.24% compared to ASIR-V MAR (paired calculation) across different dose levels and tube voltages (all p < 0.001). The noise of low-dose images with DLIR were similar to those of standard-dose images with ASIR-V, although there were statistically significant differences (p < 0.05, Table 1). Furthermore, for different tube voltages, the noise values of 70 kVp images were significantly higher than those of the other two tube voltages (p < 0.001), except for DLIR-H (p = 0.87) and DLIR-H MAR (p = 0.735) images at 0.5 mSv. The background noise values showed no statistically significant difference between MAR and non-MAR images across different tube voltages and radiation doses (Figs. 2A–2F). Under the same tube voltage and reconstruction algorithm for different radiation doses, the noise of the 3 mSv image group is lower than that of the 0.5 mSv group (all p < 0.001, File S1).

AI decreased significantly with MAR (MAR: 27.3–46.1 HU; without MAR: 60.7–115.6 HU; all p < 0.001) and with high tube voltages (except for 0.5 mSv ASIR-V MAR and DLIR-H MAR, all p < 0.001), as shown in Table 1 and Figs. 2G–2L. Compared to low-kVp images, the AI values for high-kVp images decreased except 0.5 mSv ASIR-V MAR and DLIR-H MAR (e.g., ASIR-V and 0.5 mSv 70/100/120 kVp: 113.0 [98.8, 124.0]/67.8 [61.7, 77.7]/65.3 [56.2, 78.6] HU), with the lowest AI value obtained using the high-kVp combined with the MAR algorithm (3 mSv 120 kVp in DLIR-H MAR: 27.3 [23.3, 31.4] HU; 0.5 mSv 120 kVp in DLIR-H MAR: 37.3 [28.7, 46.0] HU). Furthermore, there was no statistically significant difference in AI values between ASIR-V and DLIR-H, or between ASIR-V MAR and DLIR-H MAR , among four groups comparison (Kruskal–Wallis test among ASIR-V, ASIR-V MAR, DLIR-H and DLIR-H MAR) (Figs. 2G–2L). These results indicate that the AI of low-dose images with 120 kVp DLIR-H MAR was comparable to that of standard-dose images with 70 kVp ASIR-V MAR (p = 0.143). Under different radiation doses, the AI of images in the 3 mSv group was significantly lower than that of the 0.5 mSv group across multiple protocols (70 kVp DLIR-H, DLIR-H MAR; 100 kVp ASIR-V, DLIR-H MAR; 120 kVp ASIR-V MAR, DLIR-H, DLIR-H MAR; p < 0.001, File S2).

Figure 3 and Table 2 indicate that among the four reconstruction algorithms, only the 70 kVp 3 mSv protocol demonstrated statistically significant difference in SNR for artifact-impaired PNs (p = 0.041), whereas no significant SNR differences were observed in other comparative analyses. In the 70 kVp 3 mSv group, ASIR-V MAR showed a median SNR increase of approximately 74.5% compared to ASIR-V, while DLIR-H MAR exhibited a 137.3% higher median SNR than DLIR-H. Under the same reconstruction algorithm and radiation dose, there was no statistical difference in the SNR of artifact- impaired PNs among three tube voltages (p-values refer to Table 2). Furthermore, except for the 70 kVp ASIR-V and 70 kVp DLIR-H groups, the SNR for artifact-impaired PNs in the 3 mSv groups was significantly higher than that of the 0.5 mSv groups across all other scan protocols (with varying voltages and reconstruction algorithms), demonstrating statistical significance (p < 0.05).

Regarding the NPS, the trends of the NPS area under various kVp levels, radiation doses, and reconstruction algorithms were similar to those of background noise. Higher kVp, higher doses, or DLIR were more likely associated with the lower NPS areas. Regarding noise texture, the differences in f_peak/f_avg values between MAR and non-MAR images were insignificant, and both were closer to the reference values (3 mSv, FBP). Similarly, the influences of different voltages and radiation doses on f_peak/f_avg were also negligible in our study (Table 3 and Fig. 4).

Qualitative analysis

The interobserver agreements were significant concerning the extent of metal artifacts, the confidence in diagnosing artifact-impaired PNs, and the overall image quality (κ = 0.79, 0.82, 0.78 for 3 mSv, all p < 0.001; 0.74, 0.74, 0.73 for 0.5 mSv, all p < 0.001). The median [Q1, Q3] of the extent of metal artifacts assessments for 0.5 mSv ASIR-V, ASIR-V MAR, DLIR-H, and DLIR-H MAR were as follows: 2 [2, 2], 3 [3, 3], 2 [2, 3], and 3 [3, 3] (p < 0.001), and for 3 mSv were as follows: 2 [2, 3], 4 [4, 4], 2 [2, 3], and 4 [4, 4] (p < 0.001). The influence of artifacts in images with MAR was significantly lower than those in images without MAR, aligning with the findings from the objective evaluation. The median [Q1, Q3] of the overall image quality assessments for 0.5 mSv ASIR-V, ASIR-V MAR, DLIR-H, and DLIR-H MAR were as follows: 2 [1, 3], 3 [3, 3], 2 [2, 3], and 3 [3, 3] (p < 0.001), while for 3 mSv, the values were 2 [1, 3], 4 [3, 4], 2 [1.75, 3], and 4 [4, 4] (p < 0.001). These results indicate that DLIR-H MAR/ASIR-V MAR exhibited superior image quality compared to DLIR-H/ASIR-V (p < 0.001). In terms of diagnostic confidence for PNs, there were statistically significant differences in the subjective scores among different algorithms, the median [Q1, Q3] for 0.5 mSv ASIR-V, ASIR-V MAR, DLIR-H, and DLIR-H MAR were as follows: 1 [1, 2], 3 [2, 3], 2 [1, 2], and 4 [2.75, 4] (p < 0.001), while for 3 mSv, the values were 2 [1.75, 3], 4 [3.5, 4], 3 [3, 3], and 4 [4, 4] (p < 0.001), indicating that metal artifacts in the images significantly affected the diagnostic confidence for PNs lesions under the influence of artifacts (Table 4).