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

Jiening Zhu; Chengzhu Zhang; Jason Fan; LiCheng Kuo; Weixing Cai; Xiuxiu He; Laura Cervino; Jean Moran

eess.IVOct 27, 2026

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

Authors: Jiening Zhu, Chengzhu Zhang, Jason Fan, LiCheng Kuo, Weixing Cai, Xiuxiu He, Laura Cervino, Jean Moran,

Organizations: The authors are with the Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY 10065 USA

Abstract

Cone-beam computed tomography (CBCT) is widely used for image-guided radiotherapy, but reducing the number of projections to lower dose and scan time makes reconstruction increasingly ill-posed. We present EUReCA (End-to-end Unified REconstruction across Clinical Acquisitions), a geometry-aware feed-forward model for variable-view clinical linac CBCT. For a given acquisition trajectory, EUReCA accepts 8-100 projections with a single set of weights, without view-count-specific retraining; the same architecture and training recipe extend unchanged to other clinical trajectories. It combines learned projection features with an in-frame Feldkamp-Davis-Kress (FDK) reconstruction. The model encodes each projection with a shared multiscale backbone, back-projects features into multiresolution 3D grids using the acquisition geometry, aggregates variable numbers of views at fixed computational cost, and refines the volume with a global 3D transformer and multilevel decoder. On simulated data, EUReCA achieved 29.9-33.2 dB PSNR with 10-100 views and outperformed all applicable feed-forward baselines; it also exceeded iterative optimization-based methods while running orders of magnitude faster. On 15 measured patient CBCT scans evaluated against vendor reconstructions, EUReCA achieved the best performance at every tested view count. These results suggest that geometry-aware, measurement-anchored feed-forward reconstruction can provide a practical and flexible foundation for variable-view linac CBCT.

Figures & tables

Figure 1: Method overview. (A) Geometry-aware lifting: each of the $N$ projections is encoded by a shared pretrained backbone into a three-scale feature pyramid, and features at every scale are back-projected into 3D grids using the exact cone-beam geometry and pooled across views within angular sectors (Fig. 2 ). (B) Coarse-to-fine 3D decoder: the coarse volume is refined by a global transformer, then decoded through MLP and $\times 2$ upsampling stages with multi-scale fusion to the final volume. (C) Analytic FDK reference: an in-frame FDK reconstruction of the same projections is encoded by a 3D CNN into the matching scale pyramid and merged into the decoder at every scale through gated fusion ( $\otimes$ ).

Figure 2: Variable-view angular pooling over fixed sectors. (A) The $N$ view features extracted by the shared pretrained 2D network are organized into $K=36$ fixed angular sectors by gantry angle: views within a sector are pooled to one sector feature, each carrying its angle position code, and unobserved sectors are represented explicitly (empty token), yielding an angle-encoded sector-feature sequence whose length is independent of $N$ . (B) Per 3D location, sectors are aggregated scale-specifically: at the coarse scale a transformer self-attends across sectors through a class (CLS) token; at the mid and fine scales learned mixing weights drive a softmax-weighted aggregate. The learned aggregate is combined with mean and max statistics by a learned $1\times 1\times 1$ fusion into the feature grid.

Dataset	Region	Original	Filtered	Train	Val	Test
MELA	Thorax (mediastinum)	1,100	916	731	94	91
LUNA16	Thorax (lung)	880	611	489	60	62
RibFrac	Thorax (ribs)	660	618	503	60	55
AMOS22	Abdomen	2,383	796	654	75	67
RSNA-2023	Abdomen (trauma)	4,138	2,462	1,922	259	281
AbdomenAtlas	Abdomen	9,262	5,998	4,828	586	584

Table 3

Figure 3: Halcyon acquisition geometry. (a) The $211^{\circ}$ half-scan arc (gantry $-126^{\circ}$ to $85^{\circ}$ ), with the arc start free to rotate on the $360^{\circ}$ circle and colored markers illustrating independent sparse view draws; (b) cone-beam projection geometry with source-to-axis $\mathrm{DSO}=1000$ mm, source-to-detector $\mathrm{DSD}=1540$ mm, a $256\times 256\times 128$ reconstruction volume at $2.0$ mm isotropic voxels, and a $1.12$ mm detector pixel pitch.

Figure 4: Four additional clinical acquisitions — two orthogonal views, four fixed views, and $90^{\circ}$ / $120^{\circ}$ short arcs; markers show independent draws with the arc start free on the $360^{\circ}$ circle. Each is trained with the same architecture and recipe; only the view-selection trajectory changes at data loading.

Method	Type	10 views	20 views	50 views	100 views	Time 10/20/50/100v
Ours	FF	29.93/0.908	31.52/0.928	32.77/0.944	33.22/0.952	0.32/0.41/0.71/1.33s
ILV	FF	29.42/0.875	30.17/0.891	n/a	n/a	0.38/0.50s/OOM/OOM
DIF-Net	FF	23.91/0.724	25.19/0.746	26.71/0.786	n/a	1.73/5.00/14.9s/—
DIF-Gaussian	FF	24.39/0.741	23.30/0.611	22.76/0.560	n/a	2.06/5.64/16.9s/—
FDK	MD	15.34/0.252	17.76/0.338	21.13/0.579	22.45/0.740	0.11/0.10/0.08/0.17s
R 2 -Gaussian	MD	20.55/0.550	23.06/0.661	24.77/0.768	24.86/0.786	4.6/4.5/4.6/15.0min

Table 6

Figure 5: Qualitative comparison at 10 views (thoracic case, axial/coronal/sagittal). Top: feed-forward methods — DIF-Net, DIF-Gaussian, ILV, ours, ground truth. Bottom: measurement-driven methods — FDK, R 2 -Gaussian, IntraTomo, NAF, SAX-NeRF, ours, ground truth.

Figure 6: Qualitative comparison at 20 views (same case and layout as Fig. 5 ). Top: DIF-Net, DIF-Gaussian, ILV, ours, ground truth. Bottom: FDK, R 2 -Gaussian, IntraTomo, NAF, SAX-NeRF, ours, ground truth.

Figure 7: Qualitative comparison at 50 views (same case and layout as Fig. 5 ). Top: DIF-Net, DIF-Gaussian, ours, ground truth. Bottom: FDK, R 2 -Gaussian, IntraTomo, NAF, SAX-NeRF, ours, ground truth. ILV is not shown: it is trained per view count and no $50$ -view model was trained.

Figure 8: Qualitative comparison at 100 views , measurement-driven methods and ours (same case as Fig. 5 ): FDK, R 2 -Gaussian, IntraTomo, NAF, ours, ground truth. No feed-forward baseline supports $100$ views.

Acquisition	Views	PSNR (dB)	SSIM
Two orthogonal views	2	25.43	0.833
Four fixed views ( $0/45/90/135^{\circ}$ )	4	28.23	0.879
$90^{\circ}$ short arc	15/30/45	27.84/28.36/28.71	0.882/0.893/0.900
$120^{\circ}$ short arc	15/30/45	29.12/29.91/30.05	0.896/0.911/0.913

Table 11

Method	Type	10 views	20 views	50 views	100 views
Ours	FF	25.75 / 0.780	26.44 / 0.806	26.94 / 0.829	27.24 / 0.841
ILV	FF	25.19 / 0.725	25.62 / 0.751	OOM	OOM
DIF-Net	FF	23.57 / 0.666	23.82 / 0.679	24.72 / 0.696	n/a
DIF-Gaussian	FF	24.28 / 0.665	23.19 / 0.557	22.53 / 0.506	n/a
FDK	MD	15.70 / 0.190	17.78 / 0.259	20.70 / 0.431	22.32 / 0.582
R 2 -Gaussian	MD	22.58 / 0.427	24.36 / 0.504	25.43 / 0.611	25.64 / 0.651

Table 12

Arm	Change vs. baseline (A0)	PSNR $N{=}10/20/50$	SSIM $N{=}10/20/50$
A0	baseline: render off, gradient on, FDK on	27.90/29.52/31.05	0.8646/0.8908/0.9161
A1	+ render (DRR consistency) loss	27.83/29.43/30.88	0.8704/0.8961/0.9199
A2	$-$ multi-scale gradient (MAGA) loss	27.55/29.17/30.73	0.8664/0.8926/0.9184
A3	Zero out FDK reference (branch kept)	27.60/28.99/30.10	0.8653/0.8883/0.9085

Table 13

Figure 9: Convergence-matched data-size scaling. PSNR (a) and SSIM (b) versus number of unique training cases (log scale) at $N=10/20/50$ views, best-validation checkpoints.

Test N	$\Delta$ PSNR (dB)	95% CI
10	$+0.08$	$[-0.00,+0.17]$ n.s.
20	$+0.30$	$[+0.19,+0.41]$
50	$+0.25$	$[+0.12,+0.39]$
100	$+0.75$	$[+0.59,+0.91]$

Table 15

Cone-Beam CT

Cone-beam computed tomography (CBCT) reconstructs three-dimensional anatomy from X-ray projections for dental, interventional, and radiotherapy imaging.

See all 50 papers

RePrior: Measurement-Reconciled Prior Correction for Sparse-View CBCT Reconstruction
Oct 8, 2026Mengrui Yin, Tianxingjian Ding, Wufan Zhao +2Image Reconstruction Sparse-View CT Reconstruction
University of Central Florida · The Hong Kong University of Science and Technology (Guangzhou) · Tampere University +1
ILV: Iterative Latent Volumes for Fast and Accurate Sparse-View CT Reconstruction
Mar 16, 2026Seungryong Lee, Woojeong Baek, Joosang Lee +1Cone-Beam CT Image Reconstruction
Sungkyunkwan University · Yonsei University
EPC-RawDC: Equivariant Physics Constrained 3D Diffusion with Raw Data Consistency for CBCT to CT Image Synthesis
Oct 29, 2026Alzahra Altalib, Chunhui Li, Alessandro PerelliMedical Image Synthesis Cone-Beam CT
Faculty of Science, Engineering and Business, University of Dundee, DD1 4HN, UK. · Faculty of Applied Medical Sciences, Jordan University of Science and Technology, Irbid, 21410, Jordan · School of Cardiovascular and Metabolic Health, University of Glasgow, G12 8TA (UK).

Oct 8, 2026J/K move · Enter open · S save

RePrior: Measurement-Reconciled Prior Correction for Sparse-View CBCT Reconstruction

Mengrui Yin, Tianxingjian Ding, Wufan Zhao, Ziyang Yan+1

University of Central Florida · The Hong Kong University of Science and Technology (Guangzhou) · Tampere University+1

Image Reconstruction Sparse-View CT Reconstruction Cone-Beam CT 3D Gaussian Splatting

Open paper AskPDF arXiv

Sparse-view cone-beam computed tomography (CBCT) is severely under-constrained by limited X-ray projections. Learned reconstruction can provide complementary anatomical information, but using it as a prior in patient-specific Gaussian reconstruction introduces two challenges. First, the discrepancy between the learned prior and current reconstruction is spatially non-uniform: some local differences contain useful missing 3D information, whereas others reflect prior bias. Second, even a volumetrically beneficial correction can conflict with the observed X-ray measurements. Since X-ray measurements are line integrals, spatially separated corrections may affect overlapping rays, with their signed projection responses reinforcing or partially canceling in the measurement residual, so their feasibility cannot always be assessed independently. In this paper, we introduce RePrior, a multi-stage progressive prior-correction framework that addresses correction actionability, measurement feasibility, and representation expressivity. Selective Correction Proposal (SCP) first identifies actionable signed local corrections from prior--reconstruction discrepancies. Measurement-Constrained Reconciliation (MCR) then jointly enforces their consistency with patient-specific X-ray measurements, while Hybrid Residual Enhancement (HRE) recovers additional correction capacity through complementary voxel- and Gaussian-domain degrees of freedom around the reconciled anchor. RePrior requires neither test-time ground truth nor case-specific network fine-tuning. RePrior gains approximately 2.6~dB in 3D PSNR over R

^2

-Gaussian across LUNA16, PANORAMA, and PENGWIN. Beyond prior-assisted reconstruction, it delivers additional gains in volumetric quality and observed-view projection fidelity, with consistent gains in held-out-view fidelity in the backbone-matched evaluation.

Abstract

Cone-Beam CT

Explore similar work

RePrior: Measurement-Reconciled Prior Correction for Sparse-View CBCT Reconstruction

ILV: Iterative Latent Volumes for Fast and Accurate Sparse-View CT Reconstruction

EPC-RawDC: Equivariant Physics Constrained 3D Diffusion with Raw Data Consistency for CBCT to CT Image Synthesis