Abstract
Backgroud
Metal artifacts from neurovascular coil affect image quality in computed tomography (CT), we aim to use Physics-informed sinogram completion (PISC) to reduce metal artifacts, and compare with two traditional metal artifact reduction (MAR) methods -- Normalized MAR (NMAR), Metal Artifact Reduction for Orthopedic Implants (O-MAR) and two deep learning (DL) based methods -- convolutional neural network based metal artifact reduction (CNN-MAR) and dual domain network (DuDoNet).
Methods
40 consecutive patients who underwent endovascular coil embolization for intracranial aneurysms between July 2021 to December 2022 were included, all of whom underwent brain CT examinations. Above methods were compared quantitatively by calculating the artifact index (AI). Two blinded radiologists independently evaluated these MAR methods using a five-point scale, assessing metal artifact severity and diagnose confidence, resolution, new artifacts and the contours of different soft tissue interfaces. Friedman M test was used for quantitative and qualitative evaluation.
Results
The AI value were significantly lower in DuDoNet images when compared with FBP, NMAR, O-MAR, CNN-MAR images (p < 0.001), although no statistically significant when compared with PISC images (p = 0.181). For metal artifact severity and diagnosis confidence score, PISC method was significantly higher than FBP, NMAR, O-MAR and DuDoNet methods (all p < 0.05), although no statistically significant when compared with CNN-MAR method (p = 1.000). The resolution and contours of different soft tissue interfaces score were lower in DuDoNet images when compared with other images (all p < 0.001). The PISC method introduces the least new artifacts among these MAR methods. In addition, using traditional or DL based methods, we found new lesion obscured by metal artifacts in two cases.
Conclusions
For quantitative image analysis, DuDoNet achieved the best image quality. For qualitative image analysis, PISC achieved the best image quality. DuDoNet method can cause over-smoothing and blurring effect. PISC method introduce least new artifact.
Keywords: Metal artifact reduction, Neurovascular coil, Computed tomography, Deep learning
Background
Statistics show that approximately 3% of the population have experienced intracranial aneurysms, which are commonly treated using surgical clip or neurovascular coils, to clamp the aneurysmal neck or embolize the aneurysm [1, 2]. Given that the recovery times of endovascular coiling are faster than traditional open surgical clipping, endovascular coil embolization is the preferred minimally invasive surgical method [3]. However, the risk of new or recurrent hemorrhage (2.4%) and ischemic stroke (4.7%) remains causes for concern [4]. Follow-up examination using brain CT is essential in case of clinical deterioration. Neurovascular coils can generate metal artifacts in CT images due to beam hardening and photon starvation, which may obscure important diagnostic information and mimic abnormalities [5–8].
Mainstream metal artifact reduction methods include projection interpolation, iterative reconstruction and physical correction (PC) algorithm. Meyer et al. proposed to replace the metal-affected regions in sinogram with the forward projection of a prior image, using normalized metal artifact reduction (NMAR) [9]. O-MAR is an iterative projection modification algorithm applied clinically. Various studies reported O-MAR can efficiently reduce metal artifacts and improve the accuracy of CT Hounsfield unit (HU) measurements [10–13]. Chang et al. proposed a prior-guided algorithm that integrates the strengths of statistical methods with the benefits of projection complementation [14]. Lyu et al. introduced PDS-MAR, a new MAR algorithm for guidewire-containing CBCT that combines tubular enhancement filtering (for metal segmentation) with Delaunay triangulation (for trace inpainting) [15]. Fan et al. proposed a physics-informed nonlinear sinogram completion (PNSC) method for metal artifact correction, which consist of improved sinogram interpolation technique and a nonlinear sinogram decomposition model [16]. Physics-informed sinogram completion (PISC) [17] is a recently proposed method that combines the Normalized MAR (NMAR) algorithm [18] with physical correction techniques [19]. Recently, deep learning (DL) has demonstrated remarkable success in computational tomography tasks, including low-dose CT reconstruction and sparse-view CT reconstruction, owing to its powerful feature extraction capabilities [20–23]. An increasing number of research efforts are devoted to the development of DL-based MAR algorithms [24–27]. Zhang et al. employed a convolutional neural network (CNN) with a lightweight architecture to acquire prior images, which were used to facilitate metal projection interpolation [27]. Yu et al. proposed an algorithm that integrates a sinogram domain enhancement network into the image domain prior network, which enhanced the effectiveness of metal artifact reduction [24]. Dual domain network (DuDoNet) [28] and its enhanced version DuDoNet++ [25], utilize a hybrid domain network to learn metal projections and suppress new artifacts. Li et al. makes the first attempt at exploring Fourier convolutions for both sinogram and image domains, which can easily and cost-efficiently provide global receptive fields for sinogram restoration and image refinement [29]. Nevertheless, most deep learning-based metal artifact reduction (MAR) algorithms rely on simulated training data, whose inherent limitations in data quality may constrain their generalization performance in clinical practice. Consequently, further clinical validation is needed for MAR algorithms developed under ideal conditions. To address this issue, we systematically validated and evaluated the clinical efficacy of representative algorithms using data from 40 patients. The dataset consists exclusively of reconstructed images, with no corresponding sinograms available for analysis. To our knowledge, the O-MAR applied clinically was not compared with other MAR methods, including DL based MAR and state-of-the-art PISC method. The primary objective of our study was to determine the merits and demerits of the present algorithms concerning clinical applications and provide valuable insights to algorithm developers and doctors.
Methods
Patient cohort
The study was approved by the local institutional review board (approval number: 2023-KY-028-01), and informed consent was waived by the ethics committee due to the retrospective nature of the study. Forty-eight consecutive patients (aged > 18 years) who underwent endovascular coil embolization for intracranial aneurysms between July 2021 to December 2022 were included, all of whom underwent CT examinations. Eight patients were excluded because of motion artifacts (n = 5) and implanted other metallic materials possibly affecting the metal artifact analysis (n = 3). Finally, 40 patients (18 males, 22 females; mean age: 58 years; age range: 32–81 years) were enrolled in the study. The location of aneurysms was as follows: 22 internal carotid (IC) artery, 8 anterior communicating (A-com) artery, 1 posterior communicating (P-com) artery, 1 anterior cerebral artery, 4 posterior cerebral artery, 5 middle cerebral (MC) artery, 2 basilar artery, 1 ophthalmic artery.
CT data acquisition and processing
Non-contrast brain CT was performed using two 256 section CT (Brilliance iCT Elite and Brilliance iCT Elite FHD, Philips Healthcare). The scan protocol for Brilliance iCT Elite was as follows: tube voltage, 120 kV; tube current, 300 mAs; collimation width, 128 × 0.625 mm; gantry rotation 0.4 s; pitch, 0.925; slice thickness, 5 mm/1 mm. The CT dose index (CTDIvol) was 43 mGy, and the dose-length product (DLP) was 688.5 mGy·cm. The scanning parameters for Brilliance iCT Elite FHD were as follows: tube voltage, 120 kV; tube current, 301 mAs; collimation width, 64 × 0.625 mm; gantry rotation 0.75 s; pitch, 0.925; slice thickness, 5 mm/1 mm. The CT dose index (CTDIvol) was 43.2 mGy, and the dose-length product (DLP) was 518.1 mGy·cm.
Algorithm introduction
PISC is an advanced MAR method developed by our team. The proposed method consists of five stages: First, the metal regions are segmented from the reconstructed CT images by thresholding. Second, the NMAR algorithm is employed for interpolation to generate metal projection interpolation data Pnmar. Third, the physical correction algorithm is applied to correct the beam hardening artifact on the original metal projection to generate Ppyhs [30]. Fourth, a pixel-level adaptive weighting scheme is employed to fuse Pnmar and Ppyhs. Last, a post-processing frequency split (FS) algorithm is utilized to obtain the final refined CT image.
Quantitative image analysis
Image quality was assessed by two board-certified radiologists (with 6 and 3 years of experience in neuroimaging, respectively) through consensus assessment. On metal artifact-affected slices, four circular ROIs of approximately 40 mm2 were placed near metal coils to encompass both most obvious white and black artifacts regions. On metal artifact-free slices, four circular ROIs were placed within cerebral white matter or gray matter, carefully avoiding focal lesions and adjacent anatomical structures (Fig. 1). All ROIs defined on the original images were precisely replicated onto the NMAR, O-MAR, PISC, CNN-MAR, and DuDoNet reconstructed images to maintain identical spatial positioning across all image dataset. For quantitative analysis, the standard deviation (SD) was precisely calculated and documented within each defined ROI. Subsequently, the artifact index (AI) was computed using the following formula [31, 32], with SDcoil and SDbackground represent the averaged values obtained from four ROIs placed in metal artifact-affected slices and artifact-free slices, respectively.
Fig. 1.
Axial brain CT image show representative regions of interest (ROI) placement (circles). Four ROIs were placed around the neurovascular coil (a), and four ROIs were placed On metal artifact-free slices (b)
 |
Qualitative image analysis in patients
Image quality was independently evaluated by two board-certified neuroradiologists (with 20 and 9 years of neuroimaging experience, respectively). Both neuroradiologists were blinded to these patient data and reconstruction methods. Metal artifact extent (caused by the coil), image resolution, degree of new artifacts, and the contours of soft tissue, ventricular, sulcal, and cisternal interfaces were evaluated using Likert scale scoring (Table 1). Additionally, potential pathologies (e.g., hemorrhage and infarction) were evaluated.
Table 1.
Qualitative scoring description
| Metal artifact severity and diagnose confidence | Resolution | ||
| Score | Definition | Score | Definition |
| 1 | Marked artifacts, which completely non-diagnostic | 1 | The resolution is far lower than FBP images |
| 2 | Major artifacts, which impairing diagnosis | 2 | The resolution is lower than FBP images |
| 3 | Moderate artifacts, which easing diagnosis | 3 | The resolution of FBP images were set 3 point as control group |
| 4 | Minor artifacts, which probably easing diagnosis | 4 | The resolution is higher than FBP images |
| 5 | No artifact and no influence on diagnosis | 5 | The resolution is far higher than FBP images |
| The extent of new artifacts | The contours of different soft tissue interfaces | ||
| Score | Definition | Score | Definition |
| 1 | Absent | 1 | Completely indistinguishable |
| 2 | Mild, acceptable | 2 | Poor |
| 3 | Moderate, partially impaired diagnosis | 3 | Moderate |
| 4 | Severe, mostly impaired diagnosis | 4 | Good |
| 5 | Extensive, undiagnostic | 5 | Excellent |
Statistical analysis
All continuous variables are expressed as the mean ± standard deviation, while categorical variables were expressed in counts and percentages. The Shapiro–Wilk test was used to test the normality. Quantitative and qualitative data were compared using the Friedman M test, with post-hoc Bonferroni correction. P < 0.05 was considered to indicate statistical significance. Weighted kappa was used to analyze the inter-reader agreement. All statistical analysis were performed using SPSS software (version 27.0, release 27.0; IBM Corporation, Armonk, NY, USA) and MedCalc 20.1.4 (Ostend Belgium). The following scale for the kappa coefficient were used to indicate agreement: 0–0.20 poor consistency; 0.21–0.40 fair consistency; 0.41–0.60 moderate consistency; 0.61–0.80 good consistency; 0.81–1 excellent consistency.
Results
Quantitative image analysis
The quantitative analysis results of various metal artifact reduction methods are presented in Table 2; Fig. 2. The DuDoNet method demonstrated significantly lower AI values compared to FBP, NMAR, O-MAR, and CNN-MAR (p < 0.001), although no statistically significant when compared with PISC (p = 0.181). The AI of O-MAR, PISC, CNN-MAR images were all significantly lower than that of FBP images (all p < 0.05). The AI values of PISC and CNN-MAR images were all significantly lower than that of NMAR images (all p < 0.05). The AI values of PISC images were significantly lower than those of O-MAR images (p = 0.035).
Table 2.
Quantitative image quality
| Artifact Index (AI) | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| Descriptive Statistics | Friedman Test p |
Reconstruction method | Pairwise Comparison P | ||||||
| Mean ± SD | Score range | FBP | NMAR | O-MAR | PISC | CNN- MAR |
DuDoNet | ||
| 32.08 ± 24.16 | 1.75-124.38 | <0.001 | FBP | - | |||||
| 17.73 ± 6.25 | 8.10-30.95 | NMAR | 1.000 | - | |||||
| 14.28 ± 6.05 | 5.06–38.84 | O-MAR | 0.006 | 0.297 | - | ||||
| 10.70 ± 4.77 | 3.55–22.87 | PISC | <0.001 | <0.001 | 0.035 | - | |||
| 13.17 ± 6.77 | 4.59–31.88 | CNN-MAR | <0.001 | 0.006 | 1.000 | 0.959 | - | ||
| 8.52 ± 6.86 | 2.54–45.33 | DuDoNet | <0.001 | <0.001 | <0.001 | 0.181 | <0.001 | - | |
Data are expressed as mean±standard deviation
Fig. 2.
For neurovascular coil, The AI values of PISC images were significantly lower than FBP, NMAR and O-MAR images (all p < 0.05). The DuDoNet method created the least AI value, and which was significantly lower than that of FBP, NMAR, O-MAR and CNN-MAR images respectively (p<0.001)
Qualitative image analysis
According to the Friedman test, the PISC images exhibited the least metal artifacts and the highest diagnostic confidence among these images, although not statistically significant when compared with CNN-MAR images (p > 0.05). In contrast, the CNN-MAR method demonstrated significantly better performance than FBP, NMAR, and O-MAR methods (all p < 0.05). Additionally, the DuDoNet method performed significantly better than FBP (p = 0.028). The resolution and contours of different soft tissue interface scores of FBP, NMAR, O-MAR, PISC and CNN-MAR images were all significantly higher than that of the DuDoNet images (all p < 0.001), and there was no significance in resolution and contours of different soft tissue interface scores among the FBP, NMAR, O-MAR, PISC and CNN-MAR images (p > 0.05). Considering the FBP image as the reference standard, the PISC images demonstrated the least new artifacts. However, this was not statistically significant when compared with CNN-MAR images (p > 0.05). Qualitative analysis results for various metal artifact reduction methods are shown in Tables 3 and 4. Interobserver agreements between the two reviewers showed good-to-excellent consistency (kappa = 0.713–0.942).
Table 3.
Qualitative image quality for reviewer 1
| Images Quality Score | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Descriptive Statistics | Pairwise Comparison P | |||||||||
| A | Mean ± SD | Score range |
Friedman Test p |
Reconstruction method | FBP | NMAR | O-MAR | PISC |
CNN- MAR |
DuDoNet |
| 2.15 ± 0.58 | 1–3 | <0.001 | FBP | - | ||||||
| 2.35 ± 0.54 | 1–3 | NMAR | 1.000 | - | ||||||
| 2.52 ± 0.60 | 1–3 | O-MAR | 1.000 | 1.000 | - | |||||
| 3.58 ± 1.01 | 2–5 | PISC | <0.001 | <0.001 | <0.001 | - | ||||
| 3.25 ± 0.59 | 2–4 | CNN-MAR | <0.001 | <0.001 | 0.002 | 1.000 | - | |||
| 2.73 ± 0.64 | 1–4 | DuDoNet | 0.028 | 0.472 | 1.000 | 0.010 | 0.140 | - | ||
| B | ||||||||||
| 3.00 ± 0.00 | 3–3 | <0.001 | FBP | - | ||||||
| 3.00 ± 0.00 | 3–3 | NMAR | 1.000 | - | ||||||
| 3.00 ± 0.00 | 3–3 | O-MAR | 1.000 | 1.000 | - | |||||
| 3.00 ± 0.00 | 3–3 | PISC | 1.000 | 1.000 | 1.000 | - | ||||
| 2.93 ± 0.27 | 2–3 | CNN-MAR | 1.000 | 1.000 | 1.000 | 1.000 | - | |||
| 2.15 ± 0.36 | 2–3 | DuDoNet | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | - | ||
| C | ||||||||||
| 1.00 ± 0.00 | 1–1 | <0.001 | FBP | - | ||||||
| 2.43 ± 0.50 | 2–3 | NMAR | <0.001 | - | ||||||
| 1.83 ± 0.50 | 1–3 | O-MAR | <0.001 | 0.014 | - | |||||
| 1.28 ± 0.51 | 1–3 | PISC | 1.000 | <0.001 | 0.010 | - | ||||
| 1.45 ± 0.55 | 1–3 | CNN-MAR | 0.166 | <0.001 | 0.214 | 1.000 | - | |||
| 2.18 ± 0.45 | 1–3 | DuDoNet | <0.001 | 1.000 | 0.472 | <0.001 | <0.001 | - | ||
| D | ||||||||||
| 3.98 ± 0.16 | 3–4 | <0.001 | FBP | - | ||||||
| 3.93 ± 0.27 | 3–4 | NMAR | 1.000 | - | ||||||
| 3.93 ± 0.27 | 3–4 | O-MAR | 1.000 | 1.000 | - | |||||
| 3.93 ± 0.27 | 3–4 | PISC | 1.000 | 1.000 | 1.000 | - | ||||
| 3.93 ± 0.27 | 3–4 | CNN-MAR | 1.000 | 1.000 | 1.000 | 1.000 | - | |||
| 2.93 ± 0.35 | 2–4 | DuDoNet | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | - | ||
Qualitative score ratings and statistical results of (A) metal artifact severity and diagnostic confidence, (B) resolution, (C) The extent of new artifacts, and (D) the contours of different soft tissue interfaces
Statistically significant results are highlighted in bold in the table
SD, standard deviation
Table 4.
Qualitative image quality for reviewer 2
| Images Quality Score | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Descriptive Statistics | Pairwise Comparison P | |||||||||
| A | Mean ± SD | Score range |
Friedman Test p |
Reconstruction method | FBP | NMAR | O-MAR | PISC |
CNN- MAR |
DuDoNet |
| 2.15 ± 0.58 | 1–3 | <0.001 | FBP | - | ||||||
| 2.38 ± 0.54 | 1–3 | NMAR | 1.000 | - | ||||||
| 2.55 ± 0.64 | 1–4 | O-MAR | 0.837 | 1.000 | - | |||||
| 3.58 ± 0.96 | 2–5 | PISC | <0.001 | <0.001 | <0.001 | - | ||||
| 3.28 ± 0.60 | 2–4 | CNN-MAR | <0.001 | <0.001 | 0.002 | 1.000 | - | |||
| 2.78 ± 0.70 | 1–4 | DuDoNet | 0.017 | 0.297 | 1.000 | 0.011 | 0.197 | - | ||
| B | ||||||||||
| 3.00 ± 0.00 | 3–3 | <0.001 | FBP | - | ||||||
| 3.00 ± 0.00 | 3–3 | NMAR | 1.000 | - | ||||||
| 3.00 ± 0.00 | 3–3 | O-MAR | 1.000 | 1.000 | - | |||||
| 3.00 ± 0.00 | 3–3 | PISC | 1.000 | 1.000 | 1.000 | - | ||||
| 2.95 ± 0.22 | 2–3 | CNN-MAR | 1.000 | 1.000 | 1.000 | 1.000 | - | |||
| 2.15 ± 0.36 | 2–3 | DuDoNet | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | - | ||
| C | ||||||||||
| 1.00 ± 0.00 | 1–1 | <0.001 | FBP | - | ||||||
| 2.43 ± 0.50 | 2–3 | NMAR | <0.001 | - | ||||||
| 1.85 ± 0.48 | 1–3 | O-MAR | <0.001 | 0.028 | - | |||||
| 1.25 ± 0.44 | 1–2 | PISC | 1.000 | <0.001 | 0.002 | - | ||||
| 1.43 ± 0.50 | 1–2 | CNN-MAR | 0.214 | <0.001 | 0.075 | 1.000 | - | |||
| 2.18 ± 0.45 | 1–3 | DuDoNet | <0.001 | 1.000 | 0.837 | <0.001 | <0.001 | - | ||
| D | ||||||||||
| 4.00 ± 0.00 | 4–4 | <0.001 | FBP | - | ||||||
| 3.93 ± 0.27 | 3–4 | NMAR | 1.000 | - | ||||||
| 3.93 ± 0.27 | 3–4 | O-MAR | 1.000 | 1.000 | - | |||||
| 3.93 ± 0.27 | 3–4 | PISC | 1.000 | 1.000 | 1.000 | - | ||||
| 3.93 ± 0.27 | 3–4 | CNN-MAR | 1.000 | 1.000 | 1.000 | 1.000 | - | |||
| 2.98 ± 0.28 | 2–4 | DuDoNet | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | - | ||
Qualitative score ratings and statistical results of (A) metal artifacts severity and diagnostic confidence, (B) resolution, (C) the extent of new artifacts, and (D) the contours of different soft tissue interfaces
Statistically significant results are highlighted in bold in the table
SD, standard deviation
Discussion
Patients with aneurysms treated by neurovascular coil embolization require postoperative CT follow-up examinations. However, higher atomic number of platinum coil can produce obvious bright and dark artifacts on CT images. MAR algorithm (O-MAR, i-MAR, SEMAR) introduced by various CT vendors (Philips Healthcare, Siemens Healthineers, Canon Medical Systems Corporation) can reduce metal artifacts due to photon starvation. Kwon et al. demonstrated through patient studies and phantom trials that the O-MAR method reduces the noise and improves the accuracy of HU [11]. Katsura et al. found that the SEMAR algorithm significantly reduces metal artifacts from intracranial aneurysm coiling [33]. Fitsiori et al. reported that the iMAR algorithm significantly reduces metal artifacts caused by intracranial metallic devices. However, when combined with ADMIRE, it can produce a new artifact in the form of a photon-starvation halo [34]. Shim et al. found that O-MAR reduces metal artifacts and enhances soft-tissue visibility in CT imaging of reverse total shoulder arthroplasty [10]. However, these above-mentioned algorithms could produce new artifacts and even degrade depiction of tissue and bone. These studies only assessed vendor-specific algorithms, lacking comparison with non-commercial algorithms. Therefore, we present a novel PISC method to reduce metal coil artifacts compared with FBP, NMAR and O-MAR approaches. With the rapid development of artificial intelligence, deep learning (DL) has achieved successes in the fields of reducing metal artifacts, so we simultaneously compared DL-based methods with the above three traditional methods to reduce metal coil artifacts on CT images. Therefore, we conducted a comprehensive comparative evaluation of DL-based approaches with conventional metal artifact reduction methods (FBP, NMAR, O-MAR and PISC) for reducing coil metal artifacts in CT imaging.
The present study showed that the NMAR method produces significantly more new artifacts than other methods (Figs. 3, 4 and 5). The O-MAR method has a limited effect for reducing metal artifacts, and can cause new artifacts (Figs. 3, 4 and 5). The PISC method can significantly reduce metal artifacts, and produce few new artifacts. When the metal coils are located nears the bone, PISC has a detrimental effect for reducing metal artifacts. In contrast, DL-based MAR methods show superior efficacy, effectively compensating for the limitations observed with PISC (Fig. 5). However, the CNN-MAR method can cause blurring effect. While DuDoNet can significantly reduce metal artifacts, it can cause over-smoothing and distortion effects. Both CNN-MAR and DuDoNet methods can also produce new artifacts. In conclusion, DuDoNet algorithm achieves the best performance in quantitative assessment; Whereas, the PISC algorithm achieves the best performance in qualitative assessment. Figures 3, 4 and 5 reveal that the DuDoNet method introduces image over-smoothing effects, resulting in reduced standard deviation (SD) and artifact index (AI) values in the reconstructed images. Consequently, comprehensive evaluation demonstrated that the PISC method yielded superior image quality among these methods.
Fig. 3.
A 78-year-old woman who underwent follow-up CT two months after the right frontal lobe hemorrhage and anterior communicating artery (ACom) aneurysmal coiling. (a) Bright and dark streak metal artifacts severely obscured visualization of right frontal lobe lesion - a chronic sequela of the prior hemorrhage (arrowhead). (b) The NMAR method reduces metal artifacts close to the coil but introduces the most new artifacts than other algorithms, new artifacts can obscure the right frontal lobe lesion (arrowhead). (c) The O-MAR method reduces metal artifacts close to the coil but also introduces new artifacts, wherein the right frontal lobe lesion (arrowhead) is obscured by dark streak artifacts, which may lead to missed diagnosis. (d) The PISC method significantly reduces metal artifacts; the boundary of the right frontal lesion (arrowhead) is clear, the pons and cerebellar soft tissues are clearly displayed, and avoids introducing new artifacts. (e) The CNN-MAR method reduces metal artifacts. However, the effect of reducing metal artifacts is no more than the PISC method and degrades the density inside the metal coil (f) The DuDoNet method significantly reduces metal artifacts compared with O-MAR, but produces undesirable over-smoothing and blurring effect; the right frontal low-density lesion was removed
Fig. 4.
Axial brain CT images from a patient with giant basilar aneurysm after stent-assisted coil embolization. (a) FBP images show bright and dark streak metal artifacts around the coil, and no lesion revealed. (b, c, d, f) The NMAR, O-MAR, PISC and DuDoNet methods only reduce partial artifacts, but a portion of the lesion (arrowhead) can be found. (e) The CNN-MAR method can reduce metal artifacts and visibly reveal the lesion, which the edge is clear (arrowhead)
Fig. 5.
(a) FBP image shows the dark and bright artifact from the metal coil. (b) The NMAR method can reduces metal artifacts, but produces the most new artifacts than other methods (arrow). (c) The O-MAR method can reduce metal artifacts, but produces new artifacts (arrow). (c) The PISC method cannot reduce metal artifacts. (d) The CNN-MAR method significantly reduces metal artifacts, but produces new artifacts (arrow). (e) The DuDoNet method significantly reduces metal artifact, but produces an undesirable over-smoothing effect
In two cases, PISC and CNN-MAR images revealed new findings. The first case was a 78-year-old female patient who suffered right frontal lobe hemorrhage due to rupture of an anterior communicating artery (ACom) aneurysm and underwent follow-up CT examination two months after surgical hematoma evacuation and endovascular coiling of the ACom aneurysm. Metal artifacts obscured the hemorrhagic foci in the right frontal lobe. Metal artifacts obscured hemorrhagic foci in FBP/NMAR/O-MAR images, potentially leading to missed hypodense lesions. In contrast, PISC and CNN-MAR images can reveal chronic hemorrhagic lesion in the right frontal lobe. DuDoNet method produces undesirable over-smoothing and distortion effect that partially obscure low-density lesions (Fig. 3). The second case was a 50-year-old male patient who underwent stent-assisted coil embolization for a giant basilar artery aneurysm. In contrast to FBP images, CNN-MAR images can reveal lesion which obscured by artifacts, and the edge of lesion is clear (Fig. 4). Evaluation of preoperative CT scans identified the lesion as a giant cerebral aneurysm. In these two cases, the PISC and CNN-MAR algorithms consistently visualized more extensive lesion areas compared to FBP reconstruction, minimizing potential diagnostic errors arising from artifact confusion.
The remarkable capacity of deep learning to learn complex features has positioned it as a predominant approach in MAR research. To achieve optimal algorithm performance, developers should incorporate both clinically meaningful evaluation criteria and quantitative objective metrics in their comparative analyses. MAR algorithm development should incorporate comprehensive solutions for clinical challenges, including unpaired data and acquisition issues, to guarantee robust clinical applicability. Therefore, it would be an interesting topic to develop MAR algorithms based on unsupervised or self-supervised learning.
The study has some limitations. First, this was a retrospective study with a relatively small number of samples. Second, reducing metal coil artifacts on cerebral CT angiography was not investigated. Third, MAR methods from other vendors including Siemens, GE, and Toshiba were not included in the study; Furthermore, we only focused on neurovascular coil, and other implant types were not studied. Last, no gold standard used to verify the authenticity of images after reducing metal artifacts. Due to the different imaging mechanism of MRI, coil-related were relatively limited, and radial artifacts were not generated. In the future, traditional MRI should be performed.
Conclusions
This study demonstrates that the PISC method achieved superior metal artifact reduction performance compared to other methods, based on combined subjective and objective assessments. PISC produce less new artifacts than other methods. DL-based MAR (CNN-MAR and DuDoNet) can compensate for the deficiency of the PISC method. However, the DuDoNet method causes over-smoothing and blurring effects, leading to degraded image quality. Traditional and DL-based MAR methods provide assistance for doctors.
Acknowledgements
Not applicable.
Abbreviations
- CT
Computed tomography
- MAR
Metal artifact reduction
- NMAR
Normalized MAR
- O-MAR
Metal artifact reduction for orthopedic implants
- PISC
Physics-informed sinogram completion
- DL
Deep learning
- CNN-MAR
Convolutional neural network based metal artifact reduction
- DuDoNet
Dual domain network
- FBP
Filtered back projection
- HU
Hounsfield units
- SD
Standard deviation
- AI
Artifact index
- CTDIvol
CT dose index
- DLP
Dose-length product
- PC
Physical correction
Author contributions
ML and ZW contributed to the study conception and design. Data collection were performed by ML, YL, XY and YQ. Algorithm was developed by YG and JM. Zhi.Y and Zhu.Y contributed to the analysis of images. ML, JZ, CH and AM contributed the formal analysis. The first draft of the manuscript was written by ML, YG, JZ. JM and ZW reviewed and edited the manuscript. ML, YL, XY and CH contributed the methodology of the study. All authors have read and agreed to the published version of the manuscript.
Funding
This study was supported by funds from the National Natural Science Foundation of China (U21A6005), the Basic and Applied Basic Research Foundation of Guangdong Province (LC2016ZD028), and the President Foundation of Zhujiang Hospital, Southern Medical University (yzjj2022qn33). The funding bodies played no role in the design of the study and collection, analysis, and interpretation of data and in writing the manuscript.
Data availability
The datasets presented in this study are available from the corresponding author upon reasonable request.
Declarations
Ethics approval and consent to participate
The study was established, according to the ethical guidelines of the Helsinki Declaration, was approved by the Ethics Committee of the Zhujiang Hospital of Southern Medical University (approval number: 2023-KY-028-01).and waived the requirement for written informed consent from patients.
Consent for publication
Informed consent was waived by the ethics committee due to the retrospective nature of the study.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Mingjun Lu and Yi Guo contributed equally to this work.
Contributor Information
Jianhua Ma, Email: jhma@smu.edu.cn.
Zhibo Wen, Email: zhibowen@163.com.
References
- 1.Walcott BP, Stapleton CJ, Choudhri O, Patel AB. Flow diversion for the treatment of intracranial aneurysms. JAMA Neurol. 2016;73(8):1002–8. [DOI] [PubMed] [Google Scholar]
- 2.Molyneux AJ, Kerr RSC, Yu L-M, Clarke M, Sneade M, Yarnold JA, Sandercock P. International subarachnoid aneurysm trial (ISAT) of neurosurgical clipping versus endovascular coiling in 2143 patients with ruptured intracranial aneurysms: a randomised comparison of effects on survival, dependency, seizures, rebleeding, subgroups, and aneurysm occlusion. Lancet. 2005;366(9488):809–17. [DOI] [PubMed] [Google Scholar]
- 3.Johnston SC, Zhao S, Dudley RA, Berman MF, Gress DR. Treatment of unruptured cerebral aneurysms in California. Stroke. 2001;32(3):597–605. [DOI] [PubMed] [Google Scholar]
- 4.Runge VM, Aoki S, Bradley WGJ, Chang K-H, Essig M, Ma L, Ross JS, Valavanis A. Magnetic resonance imaging and computed tomography of the Brain—50 years of Innovation, with a focus on the future. Invest Radiol. 2015;50(9):551–6. [DOI] [PubMed] [Google Scholar]
- 5.Parizel P, Makkat S, Van Miert E, Van Goethem J, van den Hauwe L, De Schepper A. Intracranial hemorrhage: principles of CT and MRI interpretation. Eur Radiol. 2001;11(9):1770–83. [DOI] [PubMed] [Google Scholar]
- 6.Schmitt N, Wucherpfennig L, Rotkopf LT, Sawall S, Kauczor H-U, Bendszus M, Möhlenbruch MA, Schlemmer H-P, Vollherbst DF. Metal artifacts and artifact reduction of neurovascular coils in photon-counting detector CT versus energy-integrating detector CT - in vitro comparison of a standard brain imaging protocol. Eur Radiol. 2023;33(2):803–11. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Ghodasara N, Yi PH, Clark K, Fishman EK, Farshad M, Fritz J. Postoperative spinal CT: what the radiologist needs to know. Radiographics. 2019;39(6):1840–61. [DOI] [PubMed] [Google Scholar]
- 8.Gjesteby L, Man BD, Jin Y, Paganetti H, Verburg J, Giantsoudi D, Wang G. Metal artifact reduction in CT: where are we after four decades? IEEE Access. 2016;4:5826–49. [Google Scholar]
- 9.Meyer E, Raupach R, Lell M, Schmidt B, Kachelriess M. Normalized metal artifact reduction (NMAR) in computed tomography. Med Phys. 2010;37(10):5482–93. [DOI] [PubMed] [Google Scholar]
- 10.Shim E, Kang Y, Ahn JM, Lee E, Lee JW, Oh JH, Kang HS. Metal artifact reduction for orthopedic implants (O-MAR): usefulness in CT evaluation of reverse total shoulder arthroplasty. AJR Am J Roentgenol. 2017;209(4):860–6. [DOI] [PubMed] [Google Scholar]
- 11.Kwon H, Kim KS, Chun YM, Wu HG, Carlson JNK, Park JM, Kim JI. Evaluation of a commercial orthopaedic metal artefact reduction tool in radiation therapy of patients with head and neck cancer. Br J Radiol. 2015;88(1052):20140536. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Jeong S, Kim SH, Hwang EJ, Shin C-I, Han JK, Choi BI. Usefulness of a metal artifact reduction algorithm for orthopedic implants in abdominal CT: Phantom and clinical study results. AJR Am J Roentgenol. 2015;204(2):307–17. [DOI] [PubMed] [Google Scholar]
- 13.Kidoh M, Nakaura T, Nakamura S, Tokuyasu S, Osakabe H, Harada K, Yamashita Y. Reduction of dental metallic artefacts in CT: value of a newly developed algorithm for metal artefact reduction (O-MAR). Clin Radiol. 2014;69(1):e11–6. [DOI] [PubMed] [Google Scholar]
- 14.Chang Z, Ye DH, Srivastava S, Thibault J-B, Sauer K, Bouman C. Prior-Guided metal artifact reduction for iterative X-Ray computed tomography. IEEE Trans Med Imaging. 2019;38(6):1532–42. [DOI] [PubMed] [Google Scholar]
- 15.Lyu T, Wu Z, Ma G, Jiang C, Zhong X, Xi Y, Chen Y, Zhu W. PDS-MAR: a fine-grained projection-domain segmentation-based metal artifact reduction method for intraoperative CBCT images with guidewires. Phys Med Biol. 2023;68(21):215007. [DOI] [PubMed] [Google Scholar]
- 16.Fan S, Li M, Huang C, Deng X, Li H. Metal artifacts correction based on a physics-informed nonlinear sinogram completion model. Phys Med Biol. 2025;70(6). [DOI] [PubMed]
- 17.Zhu M, Zhu Q, Song Y, Guo Y, Zeng D, Bian Z, Wang Y, Ma J. Physics-informed sinogram completion for metal artifact reduction in CT imaging. Phys Med Biol. 2023;68(6). [DOI] [PubMed]
- 18.Meyer E, Raupach R, Lell M, Schmidt B, Kachelrieß M. Normalized metal artifact reduction (NMAR) in computed tomography. Med Phys. 2010;37(10):5482–93. [DOI] [PubMed] [Google Scholar]
- 19.Verburg JM, Seco J. CT metal artifact reduction method correcting for beam hardening and missing projections. Phys Med Biol. 2012;57(9):2803. [DOI] [PubMed] [Google Scholar]
- 20.Tao X, Zhang H, Wang Y, Yan G, Zeng D, Chen W, Ma J. VVBP-Tensor in the FBP algorithm: its properties and application in Low-Dose CT reconstruction. IEEE Trans Med Imaging. 2020;39(3):764–76. [DOI] [PubMed] [Google Scholar]
- 21.Hu D, Zhang Y, Liu J, Luo S, Chen Y. DIOR: deep iterative Optimization-Based Residual-Learning for Limited-Angle CT reconstruction. IEEE Trans Med Imaging. 2022;41(7):1778–90. [DOI] [PubMed] [Google Scholar]
- 22.Cha E, Chung H, Jang J, Lee J, Lee E, Ye JC. Low-Dose Sparse-View HAADF-STEM-EDX tomography of nanocrystals using unsupervised deep learning. ACS Nano. 2022;16(7):10314–26. [DOI] [PubMed] [Google Scholar]
- 23.Guo Y, Wang Y, Zhu M, Zeng D, Bian Z, Tao X, Ma J. Dual domain closed-loop learning for sparse-view CT reconstruction. In: 7th International Conference on Image Formation in X-Ray Computed Tomography: 2022: SPIE; 2022: 130–136.
- 24.Yu L, Zhang Z, Li X, Xing L. Deep sinogram completion with image prior for metal artifact reduction in CT images. IEEE Trans Med Imaging. 2021;40(1):228–38. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Lyu Y, Lin W-A, Lu J, Zhou SK. Dudonet++: encoding mask projection to reduce Ct metal artifacts. ArXiv Preprint arXiv:200100340. 2020.
- 26.Liao H, Lin W-A, Zhou SK, Luo J. ADN: artifact disentanglement network for unsupervised metal artifact reduction. IEEE Trans Med Imaging. 2020;39(3):634–43. [DOI] [PubMed] [Google Scholar]
- 27.Zhang Y, Yu H. Convolutional neural network based metal artifact reduction in X-Ray computed tomography. IEEE Trans Med Imaging. 2018;37(6):1370–81. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Lin W-A, Liao H, Peng C, Sun X, Zhang J, Luo J, Chellappa R, Zhou SK. DuDoNet: Dual domain network for CT metal artifact reduction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019;2019:10512–21.
- 29.Li Z, Gao Q, Wu Y, Niu C, Zhang J, Wang M, Wang G, Shan H. Quad-Net: Quad-Domain network for CT metal artifact reduction. IEEE Trans Med Imaging. 2024;43(5):1866–79. [DOI] [PubMed] [Google Scholar]
- 30.Verburg JM, Seco J. CT metal artifact reduction method correcting for beam hardening and missing projections. Phys Med Biol. 2012;57(9):2803–18. [DOI] [PubMed] [Google Scholar]
- 31.Wei Y, Jia F, Hou P, Zha K, Pu S, Gao J. Clinical application of multi-material artifact reduction (MMAR) technique in revolution CT to reduce metallic dental artifacts. Insights Imaging. 2020;11(1):32. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Chae H-D, Hong SH, Shin M, Choi J-Y, Yoo HJ. Combined use of virtual monochromatic images and projection-based metal artifact reduction methods in evaluation of total knee arthroplasty. Eur Radiol. 2020;30(10):5298–307. [DOI] [PubMed] [Google Scholar]
- 33.Katsura M, Sato J, Akahane M, Tajima T, Furuta T, Mori H, Abe O. Single-energy metal artifact reduction technique for reducing metallic coil artifacts on post-interventional cerebral CT and CT angiography. Neuroradiology. 2018;60(11):1141–50. [DOI] [PubMed] [Google Scholar]
- 34.Fitsiori A, Martin SP, Juillet De Saint Lager A, Gariani J, Lovblad K-O, Montet X, Vargas MI. Iterative algorithms applied to treated intracranial aneurysms. Clin Neuroradiol. 2019;29(4):741–9. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The datasets presented in this study are available from the corresponding author upon reasonable request.