Parallel Translational CT

• Parallel Translational CT is a tomographic imaging modality where the X-ray source and detector translate linearly to acquire projection data for ROI reconstructions.
• Analytical algorithms like FBP and BPF variants, and data-driven methods such as XTransCT, address challenges like truncated data and high sampling demands.
• Practical implementations reduce hardware complexity and radiation exposure, enabling real-time, cost-effective CT imaging in point-of-care and industrial applications.

Parallel Translational Computed Tomography (PTCT) is a tomographic imaging modality in which an X-ray source and detector system undergoes linear translation, rather than conventional rotation, to acquire projection data for image reconstruction. PTCT has been proposed as an ultra-low-cost, geometrically simple, and hardware-minimal alternative for region-of-interest (ROI) tomography, especially in resource-limited, intraoperative, or point-of-care environments. Recent advances in analytical algorithms, source-translation generalizations, and data-driven reconstructions have resolved longstanding issues such as truncated data artifacts, high source-sampling demands, and volumetric speed bottlenecks, making PTCT a viable modality for both clinical and industrial imaging.

1. Scanning Geometry and Data Acquisition

In PTCT, both the X-ray source and the flat-panel detector are mounted on translation tracks and move linearly in opposite directions across one or more straight segments. The object remains fixed. Projection data are acquired along these segments using a fan-beam or cone-beam geometry. The system is parameterized by: (i) global object coordinates $(x, y)$ (2D) or $(x, y, z)$ (3D); (ii) source position as a function of translation parameter $\lambda$ and the angle $\omega$ made with the $x$-axis; and (iii) detector local coordinate $t$ (for fan-beam) or $(u, v)$ (for DRR cone-beam projections).

For a source position

$$\boldsymbol\varphi(\omega, \lambda) = \bigl(-h\cos(\lambda+\omega), -h\sin(\lambda+\omega)\bigr),$$

the projection for each view $(t, \lambda, \omega)$ is

$$p(t, \lambda, \omega) = \int_{0}^{\infty} f( \boldsymbol\varphi(\omega,\lambda)+\ell \mathbf{e} ) d\ell,$$

where $(x, y, z)$0 is the unit vector from source to object point. In clinical ROI imaging, the scans are frequently truncated over the detector, yielding only local (not full-view) coverage of the object.

PTCT admits “single-translation” (1T) and “multi-translation” (2T, 3T) modes, with multiple linear passes at different angles. ROI coverage requires that all region chords be illuminated by at least two translation lines. PTCT can be further generalized to multiple-source-translation CT (mSTCT), where $(x, y, z)$1 translation lines at angles $(x, y, z)$2 provide flexible sampling and increased field of view (Wu et al., 2016, Wang et al., 2023).

2. Analytical Reconstruction Algorithms: FBP and BPF-Type

Traditional filtered backprojection (FBP) algorithms are applicable to PTCT only for non-truncated data. Severe truncation artifacts, including edge shading and streaks, arise in ROI imaging due to insufficient angular coverage (Wu et al., 2016). To address this, backprojection-filtration (BPF)-type algorithms have been developed.

The two principal BPF algorithms for PTCT ROI reconstruction are the MP-BPF and MZ-BPF methods, mathematically formulated as:

1. Backprojection of Differentiated Projections: The first step involves differentiating the projection data with respect to the view parameter $(x, y, z)$3, then backprojecting these derivatives along “linear-PI” (L-PI) lines, weighted by a geometrically motivated function $(x, y, z)$4 that compensates for ray redundancy in multi-translation setups.
2. 1D Finite Inverse Hilbert Transform: The backprojected data along each L-PI line is converted into object intensity using an explicit finite-support inverse Hilbert transform. MP-BPF adopts the You–Zeng formula; MZ-BPF uses the Manzhirov–Zeng integral.

The net effect is an exact reconstruction (within the ROI) when each ROI chord is covered by at least two translation segments, provided the weight $(x, y, z)$5 is normalized as

$(x, y, z)$6

where $(x, y, z)$7 is the number of repeated coverage rays (Wu et al., 2016).

3. Sampling, Truncation, and mSTCT Generalization

The sampling efficiency and reconstruction fidelity in PTCT are fundamentally constrained by the system’s geometry and data truncation. In standard PTCT, high-resolution imaging demands a large number of source positions (fine $(x, y, z)$8 sampling), especially if filtered along the translation parameter in V-FBP algorithms. For example, V-FBP requires $(x, y, z)$9 projections per translation to achieve high spatial resolution (Wang et al., 2023).

mSTCT extends PTCT by running several source-translation legs at different angles, enabling field-of-view (FOV) expansion and improved ROI coverage. This enables the following algorithmic developments:

• V-FBP: “Virtual projection-based filtered backprojection” regroups (λ,u) measurements into non-truncated virtual fan-beam views. High-resolution recovery requires dense sampling in $\lambda$0.
• S-BPF & D-BPF: S-BPF operates by differentiating the projections along the source translation (view) direction, while D-BPF differentiates along the detector direction. D-BPF leverages the typically much finer detector pitch $\lambda$1; spatial resolution requirements can thus be met with 75% fewer source positions compared to V-FBP, e.g., $\lambda$2 yields reconstructions comparable to V-FBP with $\lambda$3 (Wang et al., 2023).

A summary of the analytical algorithm classes for PTCT and mSTCT is provided below.

Algorithm	Differentiation Direction	Sampling Requirements
V-FBP	$\lambda$4	Dense in $\lambda$5 (high)
S-BPF	$\lambda$6	Dense in $\lambda$7
D-BPF	$\lambda$8 (detector)	Sparse in $\lambda$9

4. Data-Driven Volumetric PTCT Reconstruction

Recent research introduces deep learning-based inverse imaging for ultra-sparse PTCT, exemplified by XTransCT. This model replaces the classic CT forward model with a learned, end-to-end transformer-based architecture that reconstructs high-quality 3D volumes from only two digital reconstructed radiographs (DRRs) taken at orthogonal 45° and 135° angles (Zhang et al., 2023).

Network Structure:

• Two parallel 2D ResNet-50 backbones extract features from the DRRs.
• Features are tokenized, positionally encoded, and combined in a six-layer transformer encoder.
• Voxel-space queries are generated for centers of 32³ supervoxels, each with positional encoding.
• A six-layer transformer decoder cross-attends the projections and queries.
• Each supervoxel’s vector is passed through a 3-layer MLP to produce the final voxel outputs.

Loss and Training: Simple $\omega$0 loss is used, with no adversarial or anatomical regularization. Training is split into backbone-frozen and fine-tuning phases using Adam.

Performance: On a 50-patient clinical set, XTransCT achieves SSIM=0.77, PSNR=22.1 dB, and Dice=0.91 in 44 ms per 128³ volume. This outperforms 3D-convolutional models and naive 2D CNNs in both accuracy and speed. Cross-dataset tests on LIDC-IDRI and LNDb confirm generalizability.

5. Practical Implementation and Applications

PTCT’s linear translation geometry greatly reduces the mechanical complexity and hardware cost of the scanner. By deploying two orthogonal source-detector tracks (e.g., at 45° and 135°), volumetric information can be acquired with complementary angular coverage, enabling pose disambiguation and improved localization (Zhang et al., 2023).

Key practical characteristics include:

• ROI imaging: Focus on local regions with minimal detector length; BPF-type algorithms suppress truncation artifacts.
• Speed: Deep-learning architectures such as XTransCT enable real-time 3D reconstructions at clinically viable frame rates.
• Dose and equipment: Ultra-sparse projection sets reduce patient radiation exposure and system mechanical complexity.
• Portability and deployment: PTCT is suited for point-of-care, image-guided interventions (e.g., surgery, radiation therapy), low-cost settings, security screening, and industrial non-destructive testing (Wu et al., 2016, Zhang et al., 2023).

6. Experimental Results, Limitations, and Future Directions

Simulation and real-data experiments across multiple papers demonstrate PTCT’s viability and highlight distinct advantages and constraints:

• ROI experiments: MP-BPF and MZ-BPF reduce ROI truncation errors to ≤25 HU in 2T/3T modes, whereas FBP exhibits errors above 500 HU. Non-truncated data shows all methods converging to sub-2% RMSE when coverage suffices (Wu et al., 2016).
• Sampling savings: D-BPF permits 75% reduction in source positions with equivalent spatial resolution, verified on FORBILD and Shepp–Logan phantoms and real micro-CT data (Wang et al., 2023).
• Volumetric speed: XTransCT reconstructs a 128³ CT in 44 ms, compared to 163 ms for prior models, enabling real-time navigation and localization (Zhang et al., 2023).

Limitations:

• Multi-translation coverage is required for ROI completeness.
• BPF introduces extra computation (Hilbert inversion) and mild image blurring.
• Weights, hilbert inversion ranges, and tapers must be tuned for optimal image quality.

Future directions include:

• Extending PTCT algorithms to 3D cone-beam geometry (“virtual L-PI” lines, analogues to helical BPF).
• Data-driven approaches using attention mechanisms and voxel-wise querying.
• GPU acceleration, automatic parameter selection, and optimized weight tapers (Wu et al., 2016, Wang et al., 2023, Zhang et al., 2023).

7. Summary Table of PTCT Developments and Metrics

Method	Key Advantage	Limitation	Imaging Speed / Accuracy
FBP	Simple, analytic	Artifacts with truncated data	Fails for ROI with truncation
MP-BPF/MZ-BPF	Exact ROI, artifact-free	Requires ≥2T for full ROI	ROI RMSE ≤25 HU (truncated tests)
V-FBP	Handles truncation	Demands dense $\omega$1 sampling	Needs $\omega$2 projections
D-BPF	Sparse sampling	Weights, inversion tuning needed	75% fewer source positions; high PSNR
XTransCT	Real-time volumetric	Needs data diversity, hardware	SSIM=0.77, PSNR=22.1 dB, 44 ms/volume

PTCT leverages both analytical and learned-inverse solutions to achieve high-fidelity tomographic reconstructions from ultra-sparse translational projection data. Modern approaches enable deployment in settings where traditional rotational CT is impractical or prohibitively expensive, marking PTCT as a pivotal technique in the growing domain of point-of-care and resource-adaptive imaging (Wu et al., 2016, Wang et al., 2023, Zhang et al., 2023).