EUReCA: End-to-End Unified CBCT Reconstruction Across Clinical Acquisitions

Problem setup

Given \(N\) cone-beam projections \(\{(p_i,\theta_i)\}_{i=1}^{N}\) acquired at gantry angles \(\theta_i\), possibly non-uniformly over a full or partial arc, we seek a single feed-forward mapping \(f_\phi:\{(p_i,\theta_i)\}_{i=1}^{N}\mapsto\mathbf{V}\) to the CBCT volume \(\mathbf{V}\), with parameters \(\phi\) shared across view counts, angular distributions, and arc spans — one model for every acquisition regime seen in training, without view-count-specific retraining. The simulated geometry follows a clinical Varian Halcyon on-board imager rather than an idealised orbit.

Interactive volume viewer

Full 256×256×128 reconstructed volumes (2 mm isotropic), decoded in your browser. Choose a patient and view count, then scrub through slices with the slider or the mouse wheel over an image; adjust the display window with the presets. Volumes load on demand (1–6 MB each) and are cached once loaded. Click an image to enlarge.

Qualitative comparisons

Pick a patient (synthetic or real projections) and a view count. To keep the page light, only the middle slice of each of the three planes is shown here; the volume viewer above has full volumes. Feed-forward baselines (DIF-Net, DIF-Gaussian, ILV) are architecture-limited to ≤20 views, so at 50 and 100 views only the per-scan optimisation panel is shown. Rows: axial / coronal / sagittal.

Feed-forward methods

Per-scan optimisation methods

Reconstruction vs. number of views

One model, synthetic projections, swept from 10 to 100 views on the Halcyon 211° arc. Drag the slider to change the view count; the polar plot shows the sampled gantry angles. Display window −400 to 600 HU. Click any panel to zoom.

Patient

AbdomenAtlas · BDMAP_00002640 RSNA-2023 · 23584

Plane

Axial Coronal Sagittal

Show beside ours

Ground truth FDK Both None

Views:

10

102030405060708090100

Additional acquisition trajectories

The same architecture, training procedure and loss, trained once per trajectory with only the view-selection rule changed: two orthogonal views, four fixed gantry angles, and 90°/120° short arcs (arc start free on the 360° circle). Synthetic projections, reserved test cases. Display window −400 to 600 HU. Click any panel to zoom.

Patient

MSK prostate · 306 AbdomenAtlas · 00002874 AbdomenAtlas · 00008460

Trajectory

2 orthogonal views 4 fixed views (0/45/90/135°) 90° short arc 120° short arc

Draw

Plane

Axial Coronal Sagittal

Show beside ours

Ground truth FDK Both None

Abstract

Cone-beam computed tomography (CBCT) is widely used to guide radiotherapy delivery. Acquiring fewer projections reduces imaging dose and scan time but makes reconstruction increasingly ill-posed. We present EUReCA (End-to-end Unified REconstruction across Clinical Acquisitions), a geometry-aware feed-forward model for clinical linac CBCT. A single network reconstructs acquisitions ranging from 8 to 100 views — substantially broader than previously demonstrated by feed-forward methods. It is a unified framework which easily generalizes to diverse clinical acquisitions via only changing the training input. EUReCA encodes projections using a shared multiscale backbone and back-projects the resulting features into multiresolution 3D grids. Learned aggregators combine an arbitrary number of views at fixed computational cost at different resolution, a global 3D transformer refines the coarse representation, and a multilevel decoder recovers fine detail using both projection features and a reference Feldkamp–Davis–Kress (FDK) reconstruction. On a routine clinical patient-setup protocol, EUReCA achieves 29.9–33.2 dB PSNR with 10–100 views, outperforming all applicable feed-forward baselines. At 100 views, it also surpasses the strongest classical per-scan method in that method's well-sampled regime while reconstructing in 0.3–1.3 s. On 15 retrospectively downsampled clinical scans evaluated against vendor reconstructions, EUReCA performs best at every tested view count. These results establish geometry-aware, measurement-anchored feed-forward reconstruction as a practical approach to variable-view linac CBCT and a promising step toward a general-purpose reconstruction model.

Method

Data

Results

Simulated projections — routine patient-setup protocol

PSNR (dB) / SSIM, mean ± std over test cases. Feed-forward methods: full reserved test set (n ≈ 1328), matched train/test view counts. Per-scan optimisation methods: fixed 20-scan subset (SAX-NeRF @100v n = 14, wall-clock timeout); the "ours" row in that panel is re-scored on the same 20 scans. "—" = architecture cannot reach that view count (ILV additionally OOMs at 50 views). show ± std

Feed-forward methods

Per-scan optimisation methods

Real clinical projections — 15 Halcyon HyperSight scans

PSNR (dB) / SSIM against vendor reconstructions. At 10–100 views each method sees only 2.5–30% of the 335–407 acquired projections.

Across acquisition geometries (same model, trained per geometry)

Data scaling

Citation

@article{zhu2026eureca,
  title   = {{EUReCA}: End-to-End Unified {CBCT} Reconstruction Across Clinical Acquisitions},
  author  = {Zhu, Jiening and Zhang, Chengzhu and Fan, Jason and Kuo, LiCheng and Cai, Weixing
             and He, Xiuxiu and Cervino, Laura and Moran, Jean and Li, Xiang and Li, Tianfang and Fu, Yabo},
  journal = {arXiv preprint},
  year    = {2026}
}