1.

INTRODUCTION

Bone trauma is often accompanied by bone marrow edema (BME), which presents as elevated fluid content within the fatty yellow marrow. BME detection is conventionally performed using MRI. However, there has been an increasing interest in BME detection using x-ray CT because of its ubiquity in the emergency department, where the presence of BME might aid the diagnosis of occult fractures^1,2. Since the x-ray attenuation alone is insufficient to distinguish relative contributions of fat, water and bone to trabecular bone voxels, dual-energy (DE) imaging with three-material decomposition have been proposed for BME identifications, e.g., using a dual-source CT³. Here, we investigate the feasibility of DE BME imaging on Flat-Panel Detector (FPD) Cone-Beam CT (CBCT) with the kV-switching protocol.

The primary challenge in DE detection of BME is the narrow energy separation between fat and water-like soft-tissues. Thus, any imaging system biases need to be carefully controlled. This is particularly challenging in CBCT due to its sensitivity to x-ray scatter and detector non-idealities (e.g., glare). We propose a BME imaging framework based on a DE CBCT protocol where the kV changes every 0.5° gantry angle (corresponding to a ~50 ms switching interval). The projections are processed with fast Monte Carlo (MC)-based scatter corrections⁴, resampled onto matching view angles, and input into the two-material (fat-Aluminum) Projection-Domain Decomposition (PDD) followed by the three-material (fat-water-cortical bone) image-domain base-change^5,6. We used both simulations and CBCT benchtop experiments to evaluate impact of scatter estimation errors and other CBCT system non-idealities on the performance of the proposed DE BME imaging framework.

2.

METHOD

2.1

kV-Switching DE CBCT of the Wrist

Simulations and benchtop experiments emulated the Siemens Healthineers Multitom Rax twin robotic x-ray system in the wrist CBCT configuration shown in Fig. 1A. The source-axis distance was 750 mm, and the source-detector was 1150 mm. We investigated a realistic kV-switching DE protocol where the x-ray source alternated between the low-energy (LE) beam of 60 kV (+2 mm Al, +0.25 mm Cu) and the high-energy (HE) beam of 120 kV (+2 mm Al, +0.25 mm Cu) every 0.5° views, yielding a total 200 LE and 200 HE frames over 200° rotation.

Figure 1.

(A) The wrist CBCT scan geometry and the kV-switching protocol. (B) Thee-material (water-fat-bone) DE decomposition framework for the kV-switching CBCT.

2.3

Simulation Studies

The sensitivity of proposed framework in Fig. 1B to scatter estimation errors was investigated using a previously reported realistic x-ray system model⁹. We assumed a 400 x 400 mm2 FPD with 0.3 x 0.3 mm² pixel size and a 0.6 mm thick CsI scintillator. The x-ray source was operated at 4 mAs/frame for the 60 kV beam, and 1 mAs/frame for the 120 kV beam. X-ray spectra and linear attenuation coefficients of all materials were obtained from Spektr3.0¹⁰; quantum and detector electronic noise with correlations were included in the model. Source axial collimation was set to 150 mm (on the detector). No anti-scatter grid was installed. Ground truth scatter was simulated using the MC algorithm with a relatively large number of photon tracks (10⁷) and minimal smoothing (kernel FWHM = 4 mm).

We used two digital anthropomorphic phantoms: a normal wrist of ~50 x 75 mm² cross section shown in Fig. 2A, and a large wrist obtained by dilating the soft tissues of the normal wrist to ~80 x 105 mm² cross section. The wrists consisted of an outer fat layer (skin), an inner water region (soft tissues), and bones made of a mixture of cortical bone and fat (yellow marrow). A ~0.85 mL BME stimulus containing cortical bone and equal fractions of fat and water was placed in the radius (Fig. 2A). All mixtures fulfilled the volume conservations.

Figure 2.

(A) FDK reconstruction of digital normal wrist phantom. Material fraction maps in the BME region (green box) are shown on the right. Volume fractions are provided for each material. (B) Physical phantom consisting of a water cylinder (not shown) with four vials containing the H₂O-EtOH-K₂HPO₄ mixtures specified in the table.

2.4

Sensitivity to Scatter Correction Errors

Simulated kV-switching data were processed following the pipeline in Fig. 1B. Fast MC scatter estimation was performed using 106 photons (Fig. 3) – an order of magnitude less than the ground truth in data generation (107). We studied the sensitivity of decomposition to inaccuracies in i) shape and ii) magnitude of the estimated scatter. The shape error was introduced by adjusting FWHM of the Gaussian smoothing kernel (ranging 10 - 290 mm), which introduced an increasing distortion of the scatter as illustrated in Fig. 3. The magnitude error was generated by scaling the denoised scatter distribution (ranging 80% - 120%). A total of 9x9 combinations of errors were tested.

Figure 3.

MC-estimated LE scatter profiles using 10⁶ photon tracks without smoothing, and with 10 mm and 290 mm FWHM Gaussian kernels. Green curves show the scatter profile along the central detector row.

The performance of decompositions under scatter inaccuracy was evaluated in terms of the specificity of BME detection. We measured whether any connected water voxel clusters >1.5 mm diameter were present in the bones where there should be no fluid – i.e., outside of the true BME stimulus. To this end, the non-zero water voxels within bone boundaries were segmented; the segmentation was then processed with erosion followed by dilation using a 1.5 mm diameter sphere to remove any clusters <1.5 mm diameter. Detection specificity was computed as: V_detected – V_{detected ⋂ true}, where V_detected is the total volume of water detected (segmented) within bones, and V_{detected ⋂ true} is the volume of water detected within the true BME region.

3.

RESULTS

Fig. 4 provides a comparison of two different interframe interpolation kernels for synthesizing HE projections at view angles matching the LE data: 2D linear and cubic B-spline. (Note that in this test, the projection data was scatterfree.) The interpolations were applied onto the sinogram of each detector row separately. The differences from a reference HE sinogram simulated directly at the view angles of the LE scan are shown in Fig. 4A for the linear interpolation, and in Fig. 4B for the B- spline interpolation. The decompositions for the two synthesis techniques are displayed as the Virtual Non-Calcium images (VNCa – the attenuation-weighted sum of decomposed fat and water images^11,12) as in Fig. 4C and 4D. Linear interpolation introduced false BME signals in the ulna and metatarsals, likely due to the sharp transitions in the sinogram of these regions. B-spline interpolation, on the other hand, was free of such errors, indicating that it is an adequate choice for angular resampling of the sinogram. A reference reconstruction of the simulation with matched LE and HE frames is provided in Fig. 4E. Spline projection view synthesis was applied in the subsequent studies.

Figure 4.

Absolute sinogram errors for linear (A) and spline (B) interframe interpolations of HE data. The errors are plotted for the detector row approx. 200 slices inferior from the central slice. Green arrows indicate the error on the ulna. Decompositions (VNCa images) for DE acquisitions of unmatched LE/HE projections with linear (C) and spline (D) interframe interpolations, and of matched LE/HE projections (idealized reference) (E). Red arrows indicate the true BME region; red dashed circles mark the false BME enhancement.

Fig. 5A shows the VNCa images of the normal wrist for selected combinations of scatter magnitude and shape errors. BME (water) >1.5 mm that were detected in the bones (i.e., true and false positives) are marked. Fig. 5B and C present the BME detection specificity across investigated scatter estimation error settings for the normal and large wrist. Clear visualizations of BME and high detection specificity (<0.2 mL false detection volume compared to 0.85 mL true BME volume) was achieved for both normal and large wrists when the magnitude error was within +/-10% range. This, however, required that the scatter shape was estimated without any major distortions (smoothing FWHM <80 mm). The large wrist was more sensitive to scatter correction inaccuracies, consistent with its elevated scatter-to-primary ratio (SPRs for the normal wrist at the detector center were 6.5% and 4.6% for LE and HE frames, respectively, compared to 14.4% and 10.1% for the large wrist). Fig. 6 shows the axial and coronal VNCa images for the large wrist using scatter simulation with 10 mm FWHM smoothing and 0% introduced magnitude error. There are almost no false BME clusters outside of the true stimulus, illustrating the feasibility of DE CBCT of BME providing adequate scatter correction.

Figure 5.

(A) Axial VNCa images (cropped around the bone region) for the scatter error study of the large wrist. Red contours mark the detected BME regions (both true and false positives). (B) BME detection specificity across all scatter error cases for the normal wrist. The region in the dashed box is also shown using a narrower error range. (C) Same as (B) for the large wrist.

Figure 6.

Decomposition of the large wrist scan for MC scatter correction with 0% magnitude error and the 10 mm FWHM Gaussian smoothing kernel. Red contours mark the detected BME areas.

Fig. 7A shows the FDK reconstruction of test-bench LE acquisitions of the physical phantom. The variation in fat (i.e., EtOH) fractions among inserts (less fat – more BME) cannot be appreciated on the attenuation image, but is readily apparent in the DE decompositions with proper artifact corrections. The DE decomposition without glare or scatter correction (Fig. 7B) presents erroneous fat content in the fat-free insert (Fat-0%) and underestimated water concentrations in the water bath. These errors are largely removed with scatter correction (Fig. 7C). Fig. 7D underscores the need for the correction of FPD non-idealities – the VNCa image with glare and scatter correction achieves improved uniformity in the water bath than the image without glare correction.

Figure 7.

(A) FDK reconstruction of the test-bench LE projections of the physical BME phantom. The ground truth fat (EtOH) fraction in each insert is marked. VNCa images of decomposition with no corrections (B), without glare correction but with scatter correction (C), and with both glare and scatter corrections (D). Estimated fat fractions for each case are indicated.

4.

DISCUSSION & CONCLUSION

We investigated the feasibility of BME detection using the kV-switching DE CBCT protocol paired with an analytical two-stage three-material decomposition pipeline. The problem is inherently challenging due to unmatched x-ray paths of the LE and HE channels and similar attenuation properties of fat-like yellow marrow and water-like BME.

Using high-fidelity simulations, we found that the B-spline interpolation is sufficient to resample HE and LE frames onto common view angles required by PDD from an acquisition with a 0.5º shift between the two spectral channels. This result was achieved for the fairly complex anatomy of the wrist, indicating that spline projection synthesis might be useful in a broad range of kV-switching CBCT applications. Considering inaccuracies of scatter estimation, the decomposition maintains acceptable BME detection specificities over +/-10% magnitude errors as long as scatter shape is reasonably accurate. This finding suggests that MC or deep learning-based scatter corrections are likely adequate for BME detection using DE CBCT.

The feasibility of BME detections using proposed DE CBCT framework was validated in physical test bench experiments, demonstrating successful estimation of the difference in fluid and fat content of trabecular bone-mimicking solutions.

Ongoing works include experimental reproducibility study in complex objects under variable imaging conditions¹³, and the application of a recently reported model-based one-step DE three-material decomposition algorithm (CMBMD)⁸ to BME quantification to avoid interframe interpolations.