Physically-Grounded Manifold Projection with Foundation Priors for Metal Artifact Reduction in Dental CBCT

TL;DR

Metal artifacts in dental CBCT hinder diagnosis; the authors propose PGMP, combining Anatomically-Adaptive Physics Simulation to generate realistic polychromatic artifacts, a deterministic Direct x-Prediction restoration (DMP-Former) to project inputs onto the clean anatomical manifold in a single forward pass, and Semantic-Structural Alignment with MedDINOv3 to constrain restorations to clinically plausible anatomies. The approach yields state-of-the-art quantitative performance and robust downstream segmentation on unseen anatomy while delivering real-time inference and reduced hallucination risk. Extensive experiments on synthetic and multi-center clinical data validate superior generalization, stability, and clinical utility, supported by automated radiologist-aligned quality assessments. This framework paves the way for foundation-model-guided manifold projection in medical image restoration and could extend to other CT-based inverse problems.

Abstract

Metal artifacts in Dental CBCT severely obscure anatomical structures, hindering diagnosis. Current deep learning for Metal Artifact Reduction (MAR) faces limitations: supervised methods suffer from spectral blurring due to "regression-to-the-mean", while unsupervised ones risk structural hallucinations. Denoising Diffusion Models (DDPMs) offer realism but rely on slow, stochastic iterative sampling, unsuitable for clinical use. To resolve this, we propose the Physically-Grounded Manifold Projection (PGMP) framework. First, our Anatomically-Adaptive Physics Simulation (AAPS) pipeline synthesizes high-fidelity training pairs via Monte Carlo spectral modeling and patient-specific digital twins, bridging the synthetic-to-real gap. Second, our DMP-Former adapts the Direct x-Prediction paradigm, reformulating restoration as a deterministic manifold projection to recover clean anatomy in a single forward pass, eliminating stochastic sampling. Finally, a Semantic-Structural Alignment (SSA) module anchors the solution using priors from medical foundation models (MedDINOv3), ensuring clinical plausibility. Experiments on synthetic and multi-center clinical datasets show PGMP outperforms state-of-the-art methods on unseen anatomy, setting new benchmarks in efficiency and diagnostic reliability. Code and data: https://github.com/ricoleehduu/PGMP

Figures (13)

• Figure 1: Conceptual comparison between the stochastic noise-prediction paradigm and our deterministic manifold projection. (a) Traditional Diffusion ($\epsilon$-prediction): Iteratively removes Gaussian noise $\epsilon \sim \mathcal{N}(0, I)$ by estimating the score function in the high-dimensional ambient space. This process is computationally intensive ($T \gg 1$) and prone to accumulating stochastic errors. (b) Proposed DMP-Former ($x$-prediction): Drawing inspiration from Consistency Models, we learn a direct projection mapping $f_\theta: Y \rightarrow \mathcal{M}$. This paradigm bypasses the chaotic noise trajectory, ensuring structural determinism and efficient inference.
• Figure 2: Overview of the Hierarchical PGMP Framework. The pipeline is structured into three phases: (Top) Anatomically-Adaptive Physics Simulation (AAPS): We generate realistic training pairs by simulating polychromatic X-ray attenuation on patient-specific digital twins ($V_{metal}$), bridging the domain gap between synthetic and clinical data. (Middle) DMP-Former Network: The student network utilizes an isotropic ViT backbone with structural conditioning (AdaLN-Zero) to directly project artifact-corrupted inputs ($y$) onto the clean anatomical manifold ($x_{pred}$), bypassing iterative denoising. (Bottom) Semantic-Structural Alignment (SSA): To ensure diagnostic fidelity, a frozen medical foundation model (MedDINOv3) acts as a teacher. It guides the student network to capture expert-level anatomical semantics by minimizing the feature divergence ($\mathcal{L}_{SSA}$) between the restored and ground-truth volumes.
• Figure 3: Workflow of the proposed Anatomically-Adaptive Physics Simulation (AAPS). Step 1 & 2: Automated 3D planning inserts high-fidelity CAD models into clinically valid zones (e.g., Implant Zone $0.0h-0.60h$) to generate a volumetrically consistent metal mask $V_{metal}$. Step 3: The physics simulation incorporates a polychromatic X-ray spectrum $S(E)$ to model beam hardening according to the integral $P = -\ln \int S(E) \cdot e^{-\mu(E)L} dE$. Views: The process results in a digital twin that exhibits realistic photon starvation (Poisson noise $N \sim \text{Poisson}(\bar{N})$) and characteristic dark band artifacts.
• Figure 4: Schematic overview of the Direct Manifold Projection Transformer (DMP-Former). The network processes the artifact-corrupted volume $y$ concatenated with the structural edge mask $M_{edge}$. Unlike hierarchical U-Nets, it employs an isotropic ViT backbone with $L$ stacked DMP-Blocks. The zoom-in detail illustrates the integration of AdaLN-Zero for condition injection (regressing scale $\gamma$ and shift $\beta$), RoPE Attention for relative spatial awareness, and SwiGLU for expressive feed-forward dynamics. The model directly predicts the clean anatomy $x_{pred}$ via the $x$-prediction paradigm, minimizing the manifold reconstruction loss $\mathcal{L}_{manifold}$, while intermediate features connect to the SSA mechanism for semantic guidance.
• Figure 5: Statistical distribution analysis of the AAPS training dataset ($N=9441$). (a) Boxplot of metal cross-sectional areas across three restoration types. Implants exhibit a significantly larger area and variance ($p < 0.001$), presenting a harder restoration challenge. (b) Spatial heatmap of metal artifact centroids on the axial plane, demonstrating dense coverage of the dental arch curve without spatial bias. (c) Histogram of training slices per FDI tooth ID, confirming that the dataset includes diverse examples from both anterior (e.g., Incisors #11, #21) and posterior (e.g., Molars #16, #26) regions.