Physically-Grounded Manifold Projection Model for Generalizable Metal Artifact Reduction in Dental CBCT

🕒 Published (v1): 2025-12-30 14:36 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

PGMP (Physically-Grounded Manifold Projection) addresses metal artifact reduction in dental CBCT by combining polychromatic physics simulation for training data synthesis, a deterministic single-pass Vision Transformer (DMP-Former) that frames restoration as direct manifold projection, and a semantic alignment mechanism that uses a frozen medical foundation model (MedDINOv3) to prevent hallucinatory outputs. The framework avoids the computational burden of diffusion-based iterative sampling while preserving high-frequency anatomical detail that pixel-wise regression methods destroy.

Problem

Metal artifacts from dental implants, crowns, and fillings (beam hardening streaks, photon starvation bands) obscure trabecular bone and periodontal structures in CBCT. Supervised regression methods suffer from regression-to-the-mean (waxy, over-smoothed textures) and synthetic-to-real domain gaps from simplistic monochromatic simulation. Diffusion-based methods achieve better texture realism but require T ≫ 100 stochastic sampling steps, introduce aleatoric uncertainty incompatible with diagnostic reproducibility, and risk structural hallucination.

Method

AAPS (Anatomically-Adaptive Physics Simulation): Constructs 3D patient-specific digital twins by inserting high-fidelity CAD metal prosthetics into clinically valid anatomical zones (implants: 0.0h–0.60h, crowns: 0.40h–1.0h, fillings: 0.40h–0.95h relative to FDI tooth height). Polychromatic X-ray attenuation is modeled via energy-integrated Beer–Lambert: P = −ln ∫ S(E) exp(−∫ Σ_m x_m(r)μ_m(E)dr) dE at 120 kVp. Photon starvation is simulated with Poisson noise (N₀ = 5×10⁵ photons/bin), and scatter is approximated by a Gaussian convolution proxy (SPR = 0.1, σ = 10 px) rather than full Monte Carlo.

DMP-Former: Isotropic Vision Transformer (no hierarchical downsampling) with RoPE self-attention, AdaLN-Zero conditioning on a structural edge mask M_edge (initialized to identity to stabilize early optimization), and SwiGLU FFN. The network performs direct x-prediction—a single forward pass from corrupted input y to clean volume x_pred—grounded in Rectified Flow theory: z_t = tx + (1−t)y, with v_t = x − y constant, allowing direct manifold projection f_θ: Y → M without ODE integration.

SSA (Semantic-Structural Alignment): Frozen MedDINOv3 (pretrained on 3M+ clinical CT slices) acts as a teacher, extracting dense features F_teach from ground-truth volumes. A lightweight 3×3 conv projector maps student features to teacher dimension, preserving spatial topology. Cosine similarity loss with spatial normalization (subtract per-channel spatial mean, scale by spatial std) aligns student features to the medical semantic manifold: L_SSA = 1 − ⟨F̂′_student, F̂_teach⟩ / (‖F̂′_student‖₂ · ‖F̂_teach‖₂).

Total loss: L_total = ‖x̂ − x_gt‖₁ + 0.2·L_SSA + 0.1·L_edge (Sobel gradient L1 on tooth ROI).

Key Contributions

• AAPS pipeline with volumetrically consistent 3D digital twin placement and polychromatic spectral simulation, generating 9,441 paired training slices from the STS24 dataset
• DMP-Former: isotropic ViT with single-step deterministic manifold projection, avoiding diffusion's iterative stochastic sampling
• SSA mechanism: spatial-topology-preserving distillation from frozen MedDINOv3, replacing adversarial discriminators with a stable semantic regularizer
• Multi-center clinical evaluation protocol (multiple scanner manufacturers/protocols) with downstream segmentation validation and expert radiologist assessment

Results

Quantitative headline numbers were not present in the provided text (paper truncated before the results tables). The abstract claims PGMP "outperforms state-of-the-art baselines on unseen anatomy" and "sets a new benchmark for computational efficiency and diagnostic reliability" versus baselines including CNNMAR, DuDoNet, InDuDoNet+, RISE-MAR, and DuDoDp — all retrained on the same AAPS benchmark. No specific PSNR/SSIM/Dice figures were available in the supplied text.

Limitations

• Real-world clinical test set has no ground-truth clean images; sim-to-real transfer is assessed only qualitatively by radiologists
• Scatter model is a Gaussian convolution proxy; complex angular Compton scattering dependencies are not captured
• Evaluated exclusively on dental CBCT; generalizability to other CT modalities or body regions is undemonstrated
• MedDINOv3 was pretrained on CT slices; applicability of the SSA approach to MRI or X-ray modalities requires separate validation
• Quantitative results were truncated from the provided text, preventing independent verification of claimed performance gains

Relevance to Foundation Models in Medicine

This paper demonstrates a concrete deployment pattern for medical foundation models as frozen semantic teachers: MedDINOv3's pretrained CT representations provide anatomical priors that constrain a task-specific restoration network without fine-tuning, functioning as a substitute for adversarial discriminators. The explicit comparison of domain-adapted (MedDINOv3, DINOv3) versus natural-image-pretrained (DINOv2) encoders establishes empirical grounding for the argument that modality-specific pretraining is necessary when working with radiological density values rather than RGB photometry. The iREPA-inspired spatial alignment—preserving 2D grid topology in the distillation head—generalizes beyond MAR and is relevant to any medical restoration or reconstruction task where spatial semantics matter. For researchers tracking foundation models in medicine, this work illustrates how FM features can regularize ill-posed inverse problems in imaging without requiring large labeled datasets or task-specific FM fine-tuning.