PaperlayerView exact revision on arXiv

Source-linked AI summary

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, Xiang Li, Tianfang Li, Yabo Fu

arXiv:2610.37740v1eess.IVcs.AI

TL;DR

Reducing CBCT projections can lower dose and scan time, but makes reconstruction increasingly ill-posed, and existing feed-forward methods have not demonstrated a framework spanning clinical sampling conditions. EUReCA uses geometry-aware, measurement-anchored feed-forward reconstruction with an FDK reference. It outperforms applicable feed-forward and measurement-driven methods across sparse and well-sampled regimes while reconstructing in 0.3–1.3 s.

  • Problem

    Reducing projections makes CBCT reconstruction increasingly ill-posed, and a single feed-forward framework spanning clinical sampling conditions and trajectories has not been demonstrated.

    Source: §1 · pp. 1–2

  • Method

    EUReCA back-projects geometry-aware projection features into volumetric grids, aggregates them for global 3D refinement, and fuses them with an FDK reconstruction from the same measurements.

    Source: §1 · p. 2

  • Results

    EUReCA outperforms applicable feed-forward and measurement-driven methods across sparse and well-sampled regimes while reconstructing in 0.3–1.3 s.

    Source: §6 · p. 18

  • Takeaways & Limitations

    Reproducing a clinical acquisition geometry and testing on measured data narrows the gap between simulation and practice.

    Source: §6 · p. 18

    Evaluation should extend across institutions, machines, and protocols.

    Source: §6 · p. 19

Abstract

from arXiv · show

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.

1 Introduction

Sparse and incomplete CBCT sampling makes reconstruction ill-posed, while existing methods have not demonstrated one framework across the sampling conditions and clinical geometries of linac CBCT. EUReCA addresses this with geometry-aware reconstruction that fuses learned projection features with an FDK reference and supports flexible acquisitions.

1 Introduction

  • FDK is fast and deterministic but degrades under angular undersampling, while algebraic and regularized iterative methods require per-scan optimization and parameter tuning.Neural-field and Gaussian methods,,, remain measurement-anchored but typically take minutes to hours per volume, with weaker constraints at lower view counts.Source: §1 · pp. 1–2
  • Diffusion methods combine learned priors with projections through iterative posterior sampling, but require many network evaluations and still take far longer than one feed-forward pass.Early CBCT implementations used slice-wise priors, while later samplers reconstruct full volumes.Source: §1 · p. 2
  • Population-trained feed-forward methods progress from multiview feature-based attenuation estimates and volumetric cross-view attention to structured 3D representations and latent-volume updates,, [18], [19].Source: §1 · p. 2
  • Existing feed-forward methods have not demonstrated one framework spanning linac CBCT’s broad sampling conditions and clinically realistic trajectories.Comparisons across mismatched geometries, view counts, datasets, and metrics, and transfer to measured clinical projections, remain limited.Source: §1 · p. 2
  • EUReCA back-projects multiscale projection features into volumetric grids, aggregates them for global 3D refinement, and combines them with an FDK reconstruction from the same measurements.The full 3D FDK reference uses voxel-wise gates to adapt its contribution at each decoder scale.Source: §1 · p. 2
  • For a given trajectory, one set of weights supports 8–100 views, while the same architecture and training recipe also cover two-view, four-view, and short-arc acquisitions.The reported evaluation compares feasible methods under matched data, geometries, splits, and metrics, and includes measured patient projections against clinical reference reconstructions.Source: §1 · p. 3

2 Method

EUReCA uses a shared geometry-aware feed-forward model to reconstruct CBCT volumes from variable-view projection sequences. It combines multiscale projection features, fixed-size angular aggregation, global volumetric refinement, and a gated FDK reference, trained with intensity, structural, and gradient-alignment losses.

2.1 Overview and problem setup

  • A shared feed-forward mapping supports projection sequences with varying view counts, angular distributions, and arc spans without view-count-specific retraining.The model encodes projections, aggregates them into fixed-size 3D feature grids, refines them globally, and incorporates an in-frame FDK reconstruction at each decoder scale.Source: §2.1 · p. 3

2.2 Multi-scale projection encoder

  • A shared pretrained backbone encodes each projection into a three-scale feature pyramid for geometry-aware lifting into 3D grids.The feature pyramid is back-projected using the exact cone-beam geometry and pooled across views within angular sectors.Source: Figure 1 · p. 4

2.3 Geometry-aware back-projection and cross-view aggregation

  • Fixed K = 36 angular sectors aggregate per-voxel view features, with within-sector pooling at O(N) cost and cross-sector aggregation independent of N.The fixed-length sector sequence allows one set of weights to accept any number of views; pooling acts on back-projected features and preserves feature-grid spatial resolution.Source: §2.3 · pp. 4–5, Figure 2 · p. 5

2.4 Gated fusion of an in-frame analytic reference

  • A learned voxel-wise gate fuses multiscale projection features with an FDK reconstruction computed from the same projections.The gate uses local sampling coverage, and the FDK reference is an adaptive input rather than an exact data-consistency constraint.Source: §2.4 · p. 6

2.5 Volume refinement and decoding

  • A coarse-scale global transformer refines fused volumetric features before a coarse-to-fine decoder restores spatial detail.The decoder combines fused features across scales, and multi-scale skips restore finer projection and FDK information.Source: §2.5 · pp. 6–7

2.6 Training objective

  • Training combines tri-planar SSIM, L1 intensity error, and multiscale gradient alignment; the omitted rendering loss gave negligible improvement while more than doubling step time.The gradient-alignment term compares spatial gradients across resolutions, with fine scales emphasizing boundary alignment and coarser scales broader contrast and shape differences.Source: §2.6 · p. 7

3 Data and Experimental Setup

The study assembles and filters a large CT training set, and specifies the clinical acquisition geometry used for reconstruction.

3.1 Training data

  • 21,231 CT volumes from seven public datasets and one internal cohort were assembled, described as among the largest datasets for feed-forward CBCT reconstruction.Source: §3.1 · pp. 7–8
  • Table I summarizes CT sources before and after filtering, together with the case-level split.Source: Table I · p. 8
  • The Halcyon geometry uses a 211° half-scan and a 256 × 256 × 128 reconstruction volume with 2.0 mm isotropic voxels and 1.12 mm detector pixels.Source: Figure 3 · p. 8
  • 12,255 volumes remained after filtering for voxel spacing below 5 mm on every axis and craniocaudal extent greater than 300 mm.Source: §3.1 · p. 8

3.2 Acquisition geometry and simulated projections

Simulated projections use a clinical Halcyon geometry, with detector and patient-position modeling and several additional clinical acquisition trajectories.

3.2 Acquisition geometry and simulated projections

  • The simulation geometry reproduces the clinical Varian Halcyon HyperSight source-to-axis distance, cone angle, magnification, and half-scan trajectory.Source: §3.2 · p. 8
  • A custom CUDA cone-beam projector integrates attenuation along source-to-detector rays and renders a 384 × 768 detector with square 1.12 mm pixels.Source: §3.2 · p. 8
  • Four additional acquisitions comprise two orthogonal views, four fixed views, and 90°/120° short arcs, trained with the same architecture and recipe.Source: Figure 4 · p. 9
  • The square-pixel detector simplifies native 1.12 × 0.28 mm bins by resampling only the finer axis, while rays are integrated through CTs at original resolution.Source: §3.2 · p. 9
  • Two rigid pose augmentations apply rotations up to ±5° per axis and translations within the margin between scan extent and reconstruction field of view.Source: §3.2 · p. 9

3.3 Acquisition trajectories by sampling

The primary protocol samples variable numbers of views over a randomly started clinical half-scan; additional trajectories change view selection without changing the model or training procedure.

3.3 Acquisition trajectories by sampling

  • Training draws a new 211° half-scan acquisition with a random arc start, a log-uniformly sampled view count from 8 to 100, and a new subset of angles.Source: §3.3 · p. 9
  • The study also evaluates 90° and 120° short arcs, four-view fixed-angle acquisitions, and two orthogonal views, while keeping the architecture, loss, and training procedure unchanged.Source: §3.3 · p. 9
  • Training masks up to 7% of the detector along its upper and lower edges with probability 0.2 to model clinical beam collimation.Source: §3.3 · p. 9

3.4 Data splits

All models and baselines use the same case-level partition, and the reserved test set is kept separate from training and model selection.

3.4 Data splits

  • The 12,255 cases are split once into 9,804 training, 1,225 validation, and 1,226 test cases for every model and comparator.Source: §3.4 · p. 9
  • Case-level splitting keeps all scans from a study together, and the test set is used only for final evaluation, never training or model selection.Source: §3.4 · p. 10

3.5 Real clinical projections

The clinical evaluation uses measured CBCT projections from 15 patients across several anatomical sites and compares sparse-view reconstructions with the vendor’s clinical reconstruction reference.

3.5 Real clinical projections

  • 15 measured patient scans span head-and-neck, thoracic, and pelvic sites, testing both DRR-to-kV transfer and anatomical generalization.The sites are underrepresented in the training corpus.Source: §3.5 · p. 10
  • Metrics compare reconstructions from 10–100 projections with vendor iTools reconstructions using 335–407 projections, treated as a clinical reference rather than ground truth.Scores cover the intersecting physically valid reconstruction region.Source: §3.5 · p. 10
  • Projections receive vendor preprocessing and four adjacent detector pixels are binned to match the model’s 1.12-mm square input grid.Preprocessing includes detector-defect, scatter, beam-hardening, and spectrum corrections followed by negative-log conversion.Source: §3.5 · p. 10

3.6 Baselines and harmonized protocol

The study compares methods under harmonized data and evaluation conditions where feasible, while accounting for differences in training requirements and implementation availability.

3.6 Baselines and harmonized protocol

  • Feed-forward baselines were adapted to the reconstruction grid, HU window, filtering, sampling, augmentations, and data split, then evaluated on the full test set.The geometry-aware attenuation model was reimplemented and its projection geometry validated, but excluded after its released single-GPU pipeline failed to converge within the fixed compute budget.Source: §3.6 · p. 10
  • EUReCA uses one model trained with view counts sampled log-uniformly from [8, 100], whereas learning-based baselines are trained separately at 10, 20, 50, and 100 views where feasible.Fixed-view components or memory limits restrict some comparisons; EUReCA specialists were also trained to assess count-specific training.Source: §3.6 · p. 10
  • The measurement-driven comparison includes analytic FDK and four per-scan optimization methods, evaluated on a stratified subset of 20 scans using published optimization budgets.The optimizers require minutes to hours per volume and do not require dataset-level training.Source: §3.6 · p. 10

3.7 Evaluation metrics

Evaluation uses volumetric PSNR and tri-planar SSIM, with PSNR designated as primary; implementation details specify the training setup and reconstruction target.

3.7 Evaluation metrics

  • PSNR and tri-planar SSIM score the volumes, with PSNR primary because SSIM is part of the training loss and pixel-wise metrics can reward over-smoothing.Tri-planar SSIM averages slice-wise SSIM over axial, coronal, and sagittal orientations; all methods use one shared metric implementation.Source: §3.7 · p. 11

3.8 Implementation details

  • Training uses AdamW on two NVIDIA A100 GPUs for 50 epochs, with view counts independently sampled log-uniformly from [8, 100].Each epoch draws five sparse-view samples per training volume; gradient checkpointing and view- and voxel-chunking bound peak memory.Source: §3.8 · p. 11
  • The target is a 256 × 256 × 128 volume at 2.0-mm isotropic spacing, windowed to [−1000, 1000] HU and normalized to [0, 1].The fixed-count specialists reuse this configuration; the four-view experiment instead uses four fixed gantry angles spanning a 135° arc.Source: §3.8 · p. 11

4 Results

EUReCA is evaluated against feed-forward and measurement-driven baselines on simulated routine-protocol data, additional acquisition geometries, and measured clinical scans. It leads the reported comparisons, while the real-scan scores are referenced to vendor reconstructions rather than independent ground truth.

4.1 Comparison with state of the art on the routine patient-setup protocol

  • EUReCA achieved the highest PSNR and SSIM at every tested view count and exceeded the strongest measurement-driven method by more than 7 dB at 10 views and at least 3.6 dB at 20–100 views.It was also orders of magnitude faster; the comparison spans view counts and quality and speed, as summarized in Table II.Source: §4.1 · p. 11, Table II · p. 12
  • EUReCA reconstructs a volume in 0.3–1.3 s, two to four orders of magnitude faster than per-scan optimizers; DIF models are 5–21× slower and scale less favorably with more views.At 10 views, the qualitative comparison reports recovery of body-wall, mediastinal, and fine pulmonary-vessel detail, while neural-field optimizers appear overly smooth.Source: §4.1 · p. 12
  • As view count rises, methods converge but their ordering remains unchanged, while measurement-driven streaking recedes and feed-forward blurring of low-contrast structure persists.Panels are omitted where a method does not support the view count.Source: §4.1 · p. 12

4.2 Reconstruction across clinical acquisition geometries

  • Without architectural changes, the framework achieves 25.4–30.1 dB PSNR and 0.83–0.91 SSIM across acquisition settings; four views over 135° outperform 15 views within a 90° arc.At matched view counts, the 120° arc consistently outperforms the 90° arc; Table III reports quality across geometries using one model per geometry.Source: §4.2 · p. 12, Table III · p. 13
  • Broader angular coverage is consistent with reducing missing-wedge ambiguity that cannot be resolved by adding views within a limited arc.The geometry experiments demonstrate feasibility rather than serving as head-to-head benchmarks.Source: §4.2 · p. 13

4.3 Real-scan test: real clinical projections

  • On 15 measured clinical scans, EUReCA achieves the highest PSNR and SSIM at every view count, leading ILV by 0.56 dB at 10 views and NAF by 0.33 dB at 100 views.The real-data comparison is summarized in Table IV; each method receives 10–100 views against a vendor reconstruction made from 335–407 projections.Source: §4.3 · p. 13, Table IV · p. 14
  • The synthetic-to-real shift changes relative scores: EUReCA’s 10-view PSNR changes from 29.93 to 25.75 dB, while NAF’s changes from 21.76 to 24.32 dB.These cross-domain changes indicate differences in sensitivity to domain shift, not a direct comparison of solver quality.Source: §4.3 · pp. 14–15

5 Ablation Studies

Three controlled studies isolate model design, training-data scale, and the trade-off between one variable-view model and view-count-specific specialists, with each study varying one factor.

5.1 Study 1: leave-one-out architecture and loss factors

  • Adding the rendering loss changes PSNR by −0.11 dB and SSIM by +0.004–0.006, while increasing per-step training time by 2.1×–3.1×.The loss reprojects the predicted volume for measurement agreement; the authors omit it because it adds cost without meaningful reconstruction-quality gains.Source: §5.1 · p. 16
  • Removing MAGA lowers PSNR by approximately 0.3 dB and worsens ST-MAE by 1.4 HU at 10 views and 1.3 HU at 50 views.SSIM marginally favors the ablated model by 0.002 at each view count, consistent with its smoothing bias.Source: §5.1 · p. 16, Table V · p. 16
  • Removing the FDK reference reduces PSNR by 0.29, 0.53, and 0.95 dB at 10, 20, and 50 views, respectively.At 10 views, standalone FDK achieves only 15.34 dB, yet its inclusion still yields a measurable PSNR gain.Source: §5.1 · pp. 16–17, Table V · p. 16

5.2 Study 2: training-set-size scaling

  • Expanding training from 250 to 9,804 cases increases PSNR by 1.32, 1.74, and 1.67 dB at 10, 20, and 50 views, with no observed saturation.Log-linear fits yield 0.22–0.29 dB per doubling; SSIM begins flattening earlier than PSNR.Source: §5.2 · p. 17, Figure 9 · p. 17

5.3 Study 3: fixed-count specialists vs. the single combined model

  • At 50–100 views, specialists recover 0.25–0.75 dB; for sparse or mixed protocols, the combined model is preferable and serves every count.The combined model outperforms the corresponding specialist on 28–41% of individual cases, while four specialists require approximately four times the compute.Source: §5.3 · p. 18, Table VI · p. 18

6 Discussion and Conclusion

EUReCA combines learned anatomical priors with analytic estimates for incomplete CBCT acquisitions, with reconstructions reported in 0.3–1.3 seconds. The discussion identifies potential uses and boundaries for this approach.

6 Discussion and Conclusion

  • Near-real-time reconstruction could support repeated intrafraction imaging, while EUReCA volumes could initialize per-scan optimization; neither application was evaluated.The authors note that optimizer iterations could then enforce projection consistency rather than recover gross structure.Source: §6 · p. 18
  • EUReCA outperforms applicable feed-forward and measurement-driven methods across sparse and well-sampled regimes on simulated and real patient data, reconstructing in 0.3–1.3 seconds.The authors state that reproducing clinical acquisition geometry and evaluating measured data narrows the gap between simulation and practice.Source: §6 · p. 18
  • Scope remains limited: only angular undersampling was studied, reduced-exposure projection noise was not varied, and quantitative HU accuracy remains unverified.The authors also call for evaluation across institutions, machines, and protocols, and note that one jointly trained model across trajectories remains future work.Source: §6 · p. 19

Related papers

Loading 2610.37740v1…