DeepPriorCBCT: Prior-Enhanced CBCT Imaging

• DeepPriorCBCT is a deep learning framework that leverages explicit anatomical priors to significantly improve CBCT reconstruction from undersampled, low-dose data.
• It employs a three-stage process using VQ-VAE, latent prior modeling, and MAP estimation to ensure artifact suppression and structural fidelity.
• Applications in dental segmentation and adaptive radiotherapy demonstrate robust performance validated in large-scale multicenter trials.

DeepPriorCBCT refers to a class of deep learning frameworks that incorporate explicit anatomical or geometric priors to address the inverse problems of cone-beam computed tomography (CBCT) image reconstruction and segmentation, particularly in low-dose, sparse-view, or artifact-prone settings. Integrating learned, patient- or population-specific priors within end-to-end models, DeepPriorCBCT approaches are designed to mitigate loss of diagnostic quality due to undersampling, metallic artifacts, or poor inherent tissue contrast, and are validated on large-scale multicenter trials, head-to-head with conventional algorithms and regularization strategies.

1. Problem Formulation and Inverse-Problem Context

The CBCT inverse problem involves reconstructing a 3D attenuation map $x \in \mathbb{R}^n$ from a set of measured X-ray projections $y \in \mathbb{R}^m$ via a forward model $y = A x + \epsilon$, where $A \in \mathbb{R}^{m \times n}$ encodes the X-ray system geometry and $\epsilon$ is the measurement noise. In low-dose or sparse-view protocols—typical in thoracic or dental applications to mitigate radiation exposure—$m$ is greatly reduced (e.g., $1/6$ of standard CBCT), rendering the problem highly ill-posed. Naïve analytic reconstructions (e.g., Feldkamp-Davis-Kress [FDK]) display extensive streak artifacts and dramatic loss of structure fidelity.

Classical regularization (Tikhonov, TV, nonlocal means) is limited in stabilizing these regimes, failing notably in the presence of beam-hardening, metal artifacts, and nonunique analytic solutions arising from interior truncation or nonlinear polychromatic forward models (Park et al., 2023). DeepPriorCBCT injects rich, learned priors—capturing anatomical variability, statistical shape, or segmentation cues—directly into the reconstruction or segmentation process.

2. DeepPriorCBCT Framework: Discrete Latent Priors in CBCT Reconstruction

A canonical instantiation of DeepPriorCBCT solves the severely underdetermined reconstruction problem by explicitly learning discrete-distributed anatomical priors and fusing them with measurement fidelity in a three-stage process (Wang et al., 30 Nov 2025):

1. Neural Discrete Representation Learning: A VQ-VAE-based encoder maps full-view, high-dose FDK reconstructions $\mathrm{IF}$ to continuous features $z_e \in \mathbb{R}^{h \times w \times d \times c}$. These are quantized onto a learnable codebook $\{e_k\}$ of $K$ vectors, producing discrete latent indices $z$. A decoder reconstructs the image $\hat{x}=D(z_q)$. The VQ-VAE loss $L_{VQ}$ includes self-reconstruction error, codebook learning, and codebook-commitment penalties.
2. Latent Prior Modeling: The resulting discrete code tensor $z_i \in \{1,\dots,K\}$ is modeled with an autoregressive prior $p(z)$ (e.g., masked PixelCNN, normalizing flows), trained to maximize log-likelihood over the codegrid extracted from database IF volumes. This captures high-order anatomical structure correlations.
3. MAP Reconstruction in Latent Space: Given sparse measurements $y$, a MAP estimator seeks $z^*$ maximizing posterior belief: $\log p(z) - (1/\sigma^2)\|A D(z) - y\|_2^2$. This is solved via discrete or relaxed continuous optimization. The final CBCT volume $\hat{x} = D(z^*)$ is guaranteed to both fit the measured projections and obey the learned priors.

This approach is shown to be robust to extreme undersampling, achieving diagnostic quality from $1/6$ of the conventional dose and yielding artifact suppression and high structure fidelity superior to FDK and state-of-the-art regularized solvers.

3. Applications in Segmentation and Dental CBCT

DeepPriorCBCT also denotes segmentation frameworks tailored for artifact-robust and instance-aware CBCT analysis, especially in complex dental and radiotherapy contexts (Szczepański et al., 31 Jul 2025, Park et al., 2023, Liang et al., 2022). These typically integrate statistical shape models, semantic or regression-based loss, and geometric or registration-derived priors.

• Instance-Aware Tooth Segmentation: Statistical Shape Model (SSM) priors are learned via PCA from aligned training shapes, parameterizing anatomical variability through a low-dimensional code $\mathbf{b}$, penalized via $$E_{\mathrm{prior}}(\mathbf{b})=\mathbf{b}^T\Lambda^{-1}\mathbf{b}$$. Encoders extract shape codes and broadcast to decoders for shape-constrained segmentation. Concurrently, deep watershed regression heads predict continuous energy basins and gradients, enabling energy-based 3D instance labeling with robust boundary localization (Szczepański et al., 31 Jul 2025).
• Hybrid Segmentation-Guided Reconstruction: In dental CBCT, priors are constructed from panoramic projections and intra-oral scans (IOS), enabling robust tooth segmentation and metal artifact reduction. The network fuses FDK-reconstructed volumes with tight panoramic ROIs and IOS surface priors, enforced via $L_2$, physics-consistent, and segmentation-guided losses (Park et al., 2023). This setup allows recovery of dental morphology near metallic implants, outperforming analytic inpainting and prior adversarial-based methods.
• Deformable Registration Fusion for Radiotherapy ART: For auto-segmentation in adaptive radiotherapy, DeepPriorCBCT fuses deformably-registered (DIR) planning-CT contours as pseudo-labels and influencer volumes with DL-based 3D U-Net architectures. Influencer volumes channel prior shape/location directly into the encoding process, boosting accuracy beyond DIR-only pipelines, especially when combined with minimal expert-annotated CBCT fine-tuning (Liang et al., 2022).

4. Training Protocols, Data, and Quantitative Performance

DeepPriorCBCT frameworks consistently employ large-scale, multi-institutional datasets and rigorous cross validation, incorporating domain-specific augmentations (random crops, rotations, elastic deformations) and batch-wise training strategies.

• Sparse-View Reconstruction Trials: (Wang et al., 30 Nov 2025) includes 4,102 patients (8,675 scans) from 12 centers, with staged cohort splits for training/validation/testing. Prospective paired 138-patient cross-over trials (NCT07035977) confirm real-world equivalence to academic standards: DeepPriorCBCT achieves $\mathrm{PSNR}=28.96\pm2.87~\mathrm{dB}$, $\mathrm{SSIM}=0.79\pm0.10$, and reduces artifact energy by $>50\%$ over FDK; all while reducing dose by approximately $6\times$ ($\mathrm{DAP}=299.5~\upmu\mathrm{Gy\cdot m^2}$, $10.2~\mathrm{mGy}$) and maintaining clinical preference parity ($\kappa<0.2$).
• Dental Segmentation: (Szczepański et al., 31 Jul 2025) reports Dice coefficient $0.95\pm0.014$, Recall $0.939\pm0.032$, Hausdorff distance $1.44\pm0.70$ mm in multi-institution tests. Panoramic/IOS prior fusion (Park et al., 2023) improves SSIM ($\approx$+0.12 over LSGAN methods), reduces artifact index by $45\%$, and increases tooth Dice from $0.78\to0.91$.
• Radiotherapy Auto-Segmentation: (Liang et al., 2022) finds pure pseudo-label DL segmentation underperforms DIR (DSC $\approx$0.62 vs. $0.85$), but addition of influencer volumes raises performance to DIR levels, and fine-tuning on a small true-label set yields further gains—with up to $+0.02$ DSC on critical structures.

5. Clinical and Practical Implications

DeepPriorCBCT frameworks have demonstrated generalizability across hardware vendors, centers, and acquisition protocols. By embedding anatomical priors at the latent, feature, or segmentation levels, they enable substantial dose reduction (to $1/6$ or lower of standard protocols) without diagnostic loss. Radiologists and interventionalists rate DeepPriorCBCT-based images as equivalent to full-dose acquisitions for image-guided intervention and surgical planning (Wang et al., 30 Nov 2025).

Their impact is pronounced in workflow acceleration (fast, accurate online auto-segmentation for ART), robustness to artifacts (especially dental metal-induced), and heightened fidelity of complex anatomical structures (narrow tooth apices, head-and-neck OARs). A plausible implication is that DeepPriorCBCT reduces manual contour-editing time by over 50% in online adaptive radiotherapy settings, directly mitigating critical workflow bottlenecks (Liang et al., 2022).

6. Methodological Variants and Future Directions

While discrete latent prior frameworks represent the current state-of-the-art in DeepPriorCBCT, related approaches include adaptive experimental design for optimal sparse-view acquisition, employing deep-image-prior (DIP) regularization and information-theoretic acquisition criteria (Barbano et al., 2022). These methods, while not embedding explicit anatomical priors, share the fundamental strategy of leveraging data-driven learned regularization to guide reconstruction under dose or time constraints.

Future directions include refinement of prior models (e.g., richer autoregressive flows, patient-specific prior adaptation), real-time inference optimization, extension to 4D/temporal CBCT, and integration with multi-modal and physics-informed priors for true end-to-end Metal Artifact Reduction (MAR) and segmentation.

7. Limitations and Outlook

DeepPriorCBCT methods impose certain requirements—such as access to high-quality training cohorts, reliable intra-oral scan acquisition, or multi-algorithm deformable registration—that may limit deployment in some clinical environments. Additionally, the learning of discrete latent priors introduces architectural and optimization complexity; convergence and stability under extreme undersampling must be carefully validated. Nevertheless, large-scale, prospective, and blinded clinical trials confirm the effectiveness of DeepPriorCBCT across a variety of CBCT applications, offering a practical and scalable solution for low-dose, artifact-robust, clinically faithful image reconstruction and segmentation (Wang et al., 30 Nov 2025, Szczepański et al., 31 Jul 2025, Park et al., 2023, Liang et al., 2022).