Geometry adaptive projection-domain deep scatter estimation for multi-source semi-stationary cone-beam computed...

Cone-beam computed tomography with multi-source arrays and curved-panel detectors in stationary or semi-stationary configurations (sCBCT) has shown promise for compact, point-of-care imaging systems. Such geometries are subject to high-magnitude, x-ray scatter with complex spatial structure, and variation between source elements. This impedes the use of traditional projection-domain scatter estimators, while reconstruction artifacts hinder volume-based scatter estimators.

In this work, we propose adaptive deep scatter estimation (ADSE), an adaptive projection-domain scatter estimation technique tailored for sCBCT geometries. This technique is intended to overcome the limitations of projection- and volume-domain scatter estimators, which reduce their applicability in sCBCT configurations.

The scatter-contaminated projections from the sCBCT geometry are transformed into a view-invariant surrogate CBCT geometry. A projection-domain convolutional-neural network-based scatter estimator and scatter fluence weighting operator are applied iteratively, causing the output to converge toward accurate scatter estimates in the surrogate geometry. The final sCBCT scatter estimates are obtained by applying inverse fluence-weighting and re-transformation into the sCBCT geometry. ADSE was assessed in the projection-domain and image-domain via comparison with high-fidelity Monte Carlo (MC) simulations performed on in-silico test phantoms derived from high-quality CT scans of human heads. ADSE was compared to geometry-aware DSE approach trained directly on sCBCT data (gDSE), naïve projection-domain scatter estimation, non-iterative adaptive scatter estimation with a single fluence-weighting, and iterative Monte Carlo (iMC) scatter estimation. In the projection-domain, ADSE was evaluated via pixel value percentage error as a function of projection angle, and global mean absolute percentage error (MAPE). Residual scatter artifacts in the image domain were quantified as the voxelwise error compared to a scatter-free ground truth. A physical anthropomorphic head phantom was used for experimental validation on a sCBCT test bench integrating a curved-panel detector. Metrics included residual cupping, CT number non-uniformity, and recovery of contrast and contrast-to-noise-ratio (CNR) of thirteen embedded spherical inserts, with sizes ranging from 2 to 12 mm and nominal contrast ranging from -329 to 871 HU.

In in-silico experiments, ADSE exhibited projection-domain scatter magnitude MAPE of 3.88% for nontruncated projections, compared to iMC (MAPE = 4.42%) and gDSE (MAPE = 5.13%). However, when including truncated projections, MAPE for ADSE increased to 5.18%, while iMC and gDSE remained relatively consistent at 4.32% and 5.26%, respectively. In physical phantom experiments, the embedded spheres in uncorrected reconstructions from the test bench exhibited 58.87% contrast loss and 84.44% CNR loss, compared to an MDCT ground truth. ADSE recovered 48.67% of contrast and 25.03% of CNR, compared to 45.87% and 16.91% using iMC (gDSE, 40.45% and 21.44%). The magnitude of cupping artifacts and CT number non-uniformity decreased by 79% and 71%, respectively, compared to 85% and 53% for iMC (gDSE 114% and 59%).

We present ADSE, aiming to overcome the limitations of volume-domain and projection-domain scatter estimators that reduce their applicability in multisource sCBCT geometries with lateral truncation, angular under-sampling, and large geometrical variation between projection poses. In the presented studies, ADSE outperformed geometry-aware gDSE and volumetric iMC scatter estimation methods, resulting in reduced artifacts and improved image quality. The results illustrate the feasibility of performing scatter compensation in complex stationary CBCT geometries via a combination of conventional scatter estimation methods with geometrical warping operators tailored to the specific sCBCT geometry.