TL;DR
This paper introduces a novel image-domain algorithm, rMAR, to significantly reduce metal artifacts in low-energy virtual monochromatic images (VMIs) from CT scans. By leveraging high-energy VMIs as a "cleaner" reference, rMAR effectively preserves soft tissue detail and metal boundaries, enhancing the diagnostic utility of low-energy VMIs, which are crucial for dose reduction and contrast enhancement optimization.
What Problem Does This Paper Address?
Metal implants, common in medical imaging (e.g., hip prostheses, dental fillings, pedicle screws), create significant artifacts in CT scans. These artifacts, caused by beam hardening, scattering, and photon starvation, degrade image quality and obscure crucial anatomical details, hindering accurate diagnosis. Virtual monochromatic imaging (VMI) offers a way to mitigate these artifacts, with higher energy VMIs generally showing fewer artifacts. However, low-energy VMIs are clinically valuable for reducing contrast media dose and improving visualization of poorly enhanced structures. The challenge lies in the fact that low-energy VMIs are particularly susceptible to metal artifacts, compromising their diagnostic value. Existing metal artifact reduction (MAR) methods, while helpful, can sometimes introduce new artifacts or blur soft tissue details. This paper aims to solve the problem of severe metal artifacts in low-energy VMIs while preserving essential image quality.
Methods
The core of the proposed method, regional model-based metal artifact reduction (rMAR), operates in the image domain. The key idea is to use a high-energy VMI, which is assumed to have fewer artifacts, as a reference.
- Material Basis Image (MBI) Derivation: The process begins by deriving MBIs (specifically, water-based and bone-based images) from a sequence of VMIs acquired at different energy levels.
- Best VMI Identification: The VMI with the least artifacts is identified from the acquired VMI sequence. This is achieved by calculating the total variation (TV) of each VMI and selecting the one with the minimum TV.
- Artifact Mask Generation: An automatic masking method is employed to segment artifact-affected regions from artifact-free regions. This involves thresholding, Radon transforms, and interpolation to create an "artifact image." A sorting and thresholding segmentation (STS) technique is then used to generate a precise artifact mask.
- Regional Mapping Model: A mapping model is established in the artifact-free regions, defining the relationship between the best VMI and its corresponding MBIs. This model is typically a quadratic polynomial fit.
- Artifact Correction: This established mapping model is then applied to the artifact-affected regions of the MBIs to correct them.
- Synthesized Improved VMIs: Finally, artifact-improved MBIs are synthesized, and these are used to reconstruct new, artifact-reduced low-energy VMIs.
This approach is "regional" because it establishes the model in artifact-free areas and then applies it to correct artifact-affected regions. The method is designed to be compatible with various spectral CT systems due to its image-domain operation.
Datasets
The study utilized two types of datasets for evaluation:
- Phantom Data: A custom-designed cylindrical phantom made of solid water, containing multiple circular voids and two titanium rods. This phantom serves as a controlled environment to evaluate the algorithm's performance.
- Clinical Patient Data: Retrospective CT data from three patients with metal implants:
- A patient with pedicle screws.
- A patient with dental fillings.
- A patient with a hip prosthesis.
The paper does not mention the use of training datasets for deep learning models, as the proposed method is model-based.
Key Insights and Ideas
- Low-energy VMIs are a double-edged sword: They offer clinical benefits like dose reduction and better contrast visualization, but are severely hampered by metal artifacts.
- High-energy VMIs as a "clean" reference: The paper leverages the fact that higher energy VMIs generally exhibit fewer artifacts, using them as a basis for correcting lower energy images. This is a clever way to "borrow" clarity.
- Image-domain flexibility: Operating in the image domain makes the rMAR algorithm compatible with a wide range of spectral CT systems, increasing its practical applicability.
- Preserving anatomical detail is paramount: Unlike some MAR techniques that can blur soft tissues, rMAR aims to reduce artifacts without sacrificing clarity of surrounding structures or metal boundaries.
- Optimal VMI energy is case-dependent: The study found that the optimal VMI energy for artifact reduction varied (e.g., 100 keV for the phantom, 140 keV for some patient cases), highlighting the need for an automated selection process.
Key Results
- Qualitative Improvement: Visual inspection of phantom and patient images showed that rMAR effectively reduced metal artifacts in low-energy VMIs, particularly compared to the uncorrected VMI and the OMAR (Philips' built-in MAR) method. OMAR was observed to sometimes introduce new artifacts or oversmooth soft tissues.
- Quantitative Improvement:
- Phantom Line Profiles: rMAR-processed profiles more closely resembled the ground truth (artifact-free) profile than OMAR-processed profiles.
- ACT Number: The "Artifact Correction Threshold" (ACT) number, representing the difference between the mean CT value and a reference, was significantly lower for rMAR compared to OMAR in evaluated regions of interest (ROIs) in the phantom study. For example, rMAR showed an ACT of approximately 15 HU in ROI 1, while OMAR was around 15 HU, and the uncorrected VMI was ~85 HU.
- Combined Approach: In some complex cases (e.g., hip prosthesis), a combination of VMI + OMAR + rMAR yielded superior artifact reduction compared to either method alone.
Limitations and Future Work
- Optimal VMI Selection: The algorithm's performance relies on selecting the "best" VMI. If even high-energy VMIs suffer from severe photon starvation (e.g., in hip prosthesis imaging), the initial quality might limit the effectiveness of rMAR.
- Model Limitations: The current model lacks a direct and interpretable mapping to HU values, making direct quantitative analysis challenging for patient scans without ground truth.
- Generalizability: Future work should involve datasets from different CT scanners to assess the algorithm's generalizability.
- Multiple Implants: The robustness of rMAR with multiple metallic implants needs further demonstration.
- Polychromatic Images: Direct application to conventional polychromatic images is not possible, but the principles might be adaptable with modifications.
- 3D Extension: The current validation is on 2D axial slices; extending the framework to 3D volumetric datasets is a future direction.
- Best VMI Selection in 3D: Current best VMI selection is slice-based; incorporating adjacent slices could enhance robustness.
- Clinical Evaluation: While quantitative assessments were performed on phantoms, patient scans often lack metal-free references, necessitating qualitative assessments by radiologists. Future iterations could aim for more structured and quantifiable mappings.
Summary updated