🔬 physics

EPC-3D-Diff: Equivariant Physics Consistent Conditional 3D Latent Diffusion for CBCT to CT Synthesis

The paper proposes EPC-3D-Diff, a novel conditional 3D latent diffusion framework that integrates a physics-derived projection domain equivariance loss to synthesize high-quality, HU-accurate CT volumes from CBCT scans, significantly outperforming state-of-the-art methods in both phantom and clinical datasets.

Original authors: Alzahra Altalib, Chunhui Li, Haytham Al Ewaidat, Khaled Alawneh, Ahmad Qendel, Alessandro Perelli

Published 2026-05-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alzahra Altalib, Chunhui Li, Haytham Al Ewaidat, Khaled Alawneh, Ahmad Qendel, Alessandro Perelli

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to restore a beautiful, high-definition painting (a CT scan) from a blurry, smudged, and distorted sketch (a CBCT scan).

In the world of radiation therapy, doctors need the high-definition painting to calculate exactly how much radiation to give a patient. However, during the actual treatment, they only have the smudged sketch. The sketch is cheap and quick to take, but it's full of "noise" (like static on an old TV) and "scatter" (like light bouncing off a foggy window), making the colors (Hounsfield Units) inaccurate.

The paper introduces a new AI tool called EPC-3D-Diff to fix this sketch and turn it back into a perfect painting. Here is how it works, using simple analogies:

1. The "Smart Sketchbook" (Latent Space)

Usually, trying to fix a whole 3D volume (a whole head) at once is like trying to solve a giant 3D puzzle while blindfolded. It's too heavy and slow.

The Solution: The authors built a "smart sketchbook." First, they compress the 3D head into a smaller, simpler version (the latent space). Think of this as taking a high-resolution photo and turning it into a compact, efficient sketch that still holds all the important shapes. The AI does its heavy lifting on this compact sketch, which makes the process much faster and more stable.

2. The "Physics Detective" (Equivariant Loss)

This is the paper's biggest innovation. Most AI tools just try to make the blurry sketch look like the clear photo by guessing pixel by pixel. But sometimes, the AI gets creative and invents fake details that look good but aren't real.

The Analogy: Imagine you have a spinning top. If you rotate the top, the shadow it casts on the wall also rotates in a very specific, predictable way.
The Trick: The authors taught the AI a rule of physics: "If you rotate the 3D head, the X-ray shadows (projections) must rotate by the exact same amount."
How it helps: During training, the AI is forced to check its own work. It takes the restored head, rotates it, and checks if the resulting "shadows" match the real shadows of the target. If the AI tries to invent fake bones or wrong densities, the shadows won't match the rotation rule, and the AI gets "punished." This forces the AI to stick to the laws of physics, ensuring the result is not just pretty, but physically true.

3. The "Iterative Sculptor" (Diffusion Model)

Instead of trying to fix the image in one giant leap, the AI uses a Diffusion process.

The Analogy: Imagine a sculptor starting with a block of marble covered in noise (static). Instead of chiseling the final shape immediately, the sculptor slowly chips away the noise, step-by-step, refining the shape over many small stages until the perfect statue emerges.
The Result: This allows the AI to gradually refine the details, restoring the fine textures of the bone and soft tissue that other methods often blur out.

4. The Results: "The Magic of Mixing"

The team tested this on two different groups of data:

A Phantom: A fake head made of plastic and bone (like a mannequin) to test the basics.
Real Patients: Actual scans from two different hospitals with different scanners.

What they found:

Better Accuracy: The new method produced images that were much closer to the real, high-quality CT scans than previous methods (like CycleGAN or standard Diffusion models).
The "Mixing" Superpower: When they trained the AI on both the mannequin data and the real patient data together, it got even better. It was like teaching a chef to cook with two different sets of ingredients; the chef learned to be more adaptable and robust.
Specific Gains: On the mannequin data, the image quality improved by a huge margin (7.4 dB), and on real patient data, it improved by a solid amount (1.8 dB).

The Bottom Line

EPC-3D-Diff is a new AI that turns blurry, low-quality treatment scans into high-quality, accurate maps. It does this by:

Working in a "compressed" 3D space to be efficient.
Using a physics rule (rotation consistency) to stop the AI from hallucinating fake anatomy.
Slowly refining the image like a sculptor.

The paper claims this makes the resulting images more reliable for calculating radiation doses, helping doctors treat patients more accurately without needing an extra, expensive scan.

Technical Summary: EPC-3D-Diff

Problem Statement
In modern radiotherapy, Cone-Beam CT (CBCT) is routinely acquired for patient positioning but suffers from significant image quality degradation due to scatter, noise, beam hardening, and reconstruction artifacts. These issues lead to inaccurate Hounsfield Unit (HU) values and reduced soft-tissue contrast, rendering CBCT unsuitable for direct dose calculation and adaptive radiotherapy workflows. While synthetic CT (sCT) generation from CBCT is a critical solution, existing methods face limitations: supervised CNNs often over-smooth details; Generative Adversarial Networks (GANs) can introduce unrealistic high-frequency details and suffer from training instability; and recent conditional diffusion models, while stable, primarily operate in the image domain without explicitly enforcing measurement consistency with the underlying acquisition physics. Furthermore, these models often struggle to generalize across different scanner geometries and domain shifts.

Methodology
The authors propose EPC-3D-Diff (Equivariant Physics Consistent Conditional 3D Latent Diffusion), a framework designed to synthesize high-fidelity CT volumes from CBCT inputs by integrating acquisition physics directly into the diffusion process.

3D Latent Space Diffusion: To handle volumetric data efficiently, the method operates in a compact latent space learned by a lightweight 3D autoencoder (AE). The AE downsamples in-plane resolution while preserving axial depth. A conditional 3D U-Net predicts noise in this latent space, conditioned on the encoded CBCT latent representation ( $z_c$ ), following a variance-preserving diffusion formulation.
Projection Domain Equivariance: The core innovation is a physics-consistent equivariance loss derived from the CT forward model. The framework exploits the physical property that an in-plane rotation of a volume corresponds to an angular shift in its projections.
- During training, the synthesized CT volume ( $\hat{x}_0$ ) is rotated ( $R_\phi$ ) and forward-projected using the CT operator ( $A_0$ ).
- This is compared against the ground-truth CT projections ( $y_0$ ) that have been angularly shifted ( $T_\phi$ ).
- The loss ( $L_{eq}$ ) minimizes the $L_2$ -norm between these projection domains: $\sum \|T_\phi(y_0) - A_0 R_\phi(\hat{x}_0)\|^2$ .
- Theoretically, this acts as a physics-informed prior, penalizing solutions that violate acquisition symmetry and projecting intermediate diffusion samples onto an equivariant sub-manifold.
Composite Training Objective: The total loss combines the standard diffusion objective ( $L_{DDPM}$ ) with the projection equivariance loss ( $L_{eq}$ ) and complementary image-domain losses: $L_1$ for voxel accuracy, gradient loss ( $L_{edge}$ ) for structural edges, and Laplacian loss ( $L_{lap}$ ) for high-frequency details.
Inference: At test time, the model performs operator-free inference using deterministic DDIM updates. It requires only the CBCT input, making it computationally efficient for clinical deployment.

Key Contributions

Physics-Consistent Equivariance: The paper introduces a novel loss function that enforces rotational equivariance in the projection domain, directly linking the generative model to the physics of CT acquisition (Beer's law and forward operators) rather than just image statistics.
3D Latent Conditional Diffusion: The framework combines scalable 3D volumetric synthesis via latent diffusion with the aforementioned physics constraints, addressing the computational challenges of 3D medical imaging.
Robustness to Domain Shifts: By integrating acquisition physics as an inductive bias, the model is designed to generalize better across different scanner geometries and mixed-domain training scenarios compared to standard image-domain diffusion models.

Experimental Results
The method was validated on two head and neck datasets: the NWH phantom dataset (10 patients) and the JUST clinical dataset (14 patients). Evaluations included single-domain and mixed-domain (multi-scanner) training regimes.

Quantitative Performance: EPC-3D-Diff significantly outperformed state-of-the-art baselines (CycleGAN and a baseline conditional diffusion model, C-DDPM).
- Phantom Data: Achieved a +7.4 dB improvement in PSNR over C-DDPM in mixed-domain training (38.44 dB vs. 31.00 dB) and an SSIM of 0.99.
- Clinical Data: Achieved a +1.8 dB improvement in PSNR over C-DDPM.
- The method also demonstrated superior HU accuracy within tissue boundaries and lower variability compared to baselines.
Qualitative Performance: Visual comparisons showed sharper anatomical boundaries and fewer HU-dependent distortions, particularly in heterogeneous regions.
Efficiency: While training incurs a minor overhead (less than 15% of total time) due to periodic forward projections, inference is fast (58 seconds per patient) as it requires no projection operations.

Significance
The paper claims that EPC-3D-Diff improves the robustness and physics consistency of CBCT-to-CT synthesis, which is essential for downstream radiotherapy workflows requiring accurate HU values for dose calculation. By embedding acquisition physics as a structural inductive bias, the approach addresses the limitations of purely data-driven methods, offering a solution that generalizes well under limited training data and domain shifts without requiring complex operator-based corrections during inference.