⚡ electrical engineering

Cone-Beam CT Image Quality Enhancement Using A Latent Diffusion Model Trained with Simulated CBCT Artifacts

This paper proposes a self-supervised latent diffusion model trained on spatially consistent pseudo-CBCT images to enhance Cone-Beam CT image quality while effectively preserving anatomical structures and avoiding the overcorrection issues common in conventional methods.

Original authors: Naruki Murahashi, Mitsuhiro Nakamura, Megumi Nakao

Published 2026-03-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Naruki Murahashi, Mitsuhiro Nakamura, Megumi Nakao

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 navigate a city using a map. Sometimes, you have a high-definition, crystal-clear map (this is a standard CT scan). Other times, you only have a fuzzy, smudged, and low-contrast photocopy of that map (this is a Cone-Beam CT or CBCT).

Doctors use the fuzzy CBCT maps during radiation therapy to make sure they are hitting the right target (like a tumor). But because the CBCT is so blurry and full of "static" (artifacts), it's hard to see the details. If they try to sharpen the image using old computer tricks, the computer often gets confused and accidentally redraws the city streets, moving buildings or erasing parks. In medical terms, this means the computer might accidentally change the shape of a patient's organs, which is dangerous.

This paper introduces a new, smarter way to fix the blurry map without accidentally moving the buildings. Here is how they did it, broken down into simple concepts:

1. The Problem: The "Real World" vs. The "Training Class"

To teach a computer to fix blurry images, you usually need to show it a "Before" (blurry) and an "After" (clear) picture of the exact same thing.

The Catch: In the real world, a patient's body moves between the time you take the clear scan and the blurry scan. Their organs shift, they breathe, and they change position.
The Old Mistake: If you teach the computer using these mismatched real-life pairs, the computer learns to "fix" the blur by also "fixing" the movement. It might think a lung that moved is an error and try to push it back, effectively changing the patient's anatomy. This is called overcorrection.

2. The Solution: The "Perfect Fake" (Pseudo-CBCT)

Instead of trying to find perfect real-life pairs (which is impossible), the researchers decided to create their own perfect training data.

The Analogy: Imagine you have a perfect, high-definition photo of a city. Instead of waiting for a real blurry photo, you take your perfect photo and run it through a "Blur Filter" app that you control. You add the exact same smudges, streaks, and low contrast that real CBCT machines produce.
Why this works: Because you started with the perfect photo and made the blurry one, you know exactly what the "After" picture should look like. They are perfectly aligned. The computer learns to remove the smudges without ever having to guess about moving organs, because the organs never moved in the first place.

3. The Engine: The "Latent Diffusion Model"

The researchers used a type of AI called a Diffusion Model.

The Analogy: Think of a noisy, static-filled TV screen. A diffusion model is like a smart noise-canceling headphone for images. It starts with pure static and slowly, step-by-step, removes the noise until a clear image emerges.
The "Latent" Twist: Usually, this process is very heavy on computer power, like trying to clean a whole city street by hand. The researchers put the image into a "compressed suitcase" (called Latent Space) first. They clean the compressed version, then unpack it. This makes the process much faster and cheaper, like cleaning a miniature model of the city instead of the real thing.

4. The Results: A Perfect Restoration

They tested this on 75 patients with prostate cancer.

The Old Way: When using real data, the computer changed the anatomy (moved organs) in thousands of pixels. It was like the computer redrawing the map and accidentally moving a hospital to a different street.
The New Way: Their method changed the anatomy in less than 1/1000th of the pixels. It was like cleaning the smudges off the map without moving a single building.
Speed: Because they used the "compressed suitcase" method, the computer worked much faster, making it practical for real hospitals.

The Bottom Line

This paper presents a clever trick: Don't try to learn from messy, moving real-life data. Instead, create a perfect, controlled "fake" version of the problem to train the AI.

By doing this, they built an AI that can turn a fuzzy, low-quality medical scan into a sharp, high-quality one without accidentally reshaping the patient's body. It's like having a restoration artist who can clean a muddy painting perfectly, knowing exactly what the original colors were, without ever accidentally painting over the original brushstrokes.

1. Problem Statement

Cone-Beam Computed Tomography (CBCT) is widely used in radiotherapy for patient positioning but suffers from significant image quality issues compared to conventional CT, including low contrast, high noise, and various artifacts (streaking, beam-hardening, scatter).

The Core Challenge: Improving CBCT image quality using deep learning typically requires paired training data (CBCT and corresponding high-quality CT). However, acquiring perfectly aligned pairs from the same patient is difficult because:
1. CBCT and planning CT are taken at different times.
2. Patient posture, respiration, and organ motion cause anatomical deformations (especially in the trunk/pelvis).
The Consequence: Standard supervised learning (e.g., conditional diffusion models) or unsupervised methods (e.g., CycleGAN) often lead to "overcorrection." The model may alter anatomical structures (e.g., changing organ shapes, creating new air pockets, or distorting bone geometry) while attempting to fix artifacts, which compromises the accuracy of dose calculations and treatment planning.

2. Methodology

The authors propose a Conditional Latent Diffusion Model (CLDM) trained on pseudo-CBCT images generated via self-supervised learning.

A. Pseudo-CBCT Image Creation (Data Generation)

Instead of using real, misaligned CT-CBCT pairs, the authors generate "pseudo-CBCT" images directly from high-quality CT images by simulating CBCT artifacts. This ensures perfect spatial correspondence between the input (pseudo-CBCT) and the target (original CT).

Process:
1. Sinogram Domain: Apply Radon transform to the CT image. Simulate motion and scatter by warping the sinogram using a random displacement field and clipping bone portions to prevent excessive artifacts.
2. Contrast Adjustment: Apply contrast degradation parameters to the warped sinogram to simulate beam-hardening and scatter.
3. Image Domain: Reconstruct the image via inverse Radon transform. Apply Gamma correction and pixel value shifts to specific masks (center, periphery, and edges) to simulate geometric factors and field-of-view limitations.
Result: A dataset of perfectly aligned CT (Target) and Pseudo-CBCT (Condition) pairs.

B. Conditional Latent Diffusion Model (CLDM)

To address the high computational cost of standard diffusion models on high-resolution medical images, the authors extend the framework to latent space.

Architecture:
- Encoder/Decoder: Uses a pre-trained VQVAE (Vector Quantized Variational Autoencoder) to compress images into a latent space.
- Diffusion Process: A U-Net based diffusion model operates on the latent variables. It learns to predict noise added to the target latent variable ( $z_0$ ) conditioned on the pseudo-CBCT latent variable ( $x$ ).
- Training: The model minimizes the squared error between predicted noise and actual noise.
Inference: The model takes a real CBCT image (encoded to latent space), adds Gaussian noise, and iteratively denoises it conditioned on the input to generate a high-quality "Synthetic CT" (SynCT).
3D Consistency: To prevent inter-slice discontinuities, the same initial noise is used for all slices during generation.

3. Key Contributions

Overcorrection-Free Translation: By using spatially consistent pseudo-data, the method achieves image-to-image translation without altering anatomical structures, solving the "hallucination" problem common in trunk CBCT enhancement.
Latent Space Efficiency: Extending the conditional diffusion model to latent space significantly reduces computation time and memory requirements, making it feasible for high-resolution medical imaging under constrained resources.
Simulated Artifact Training: Demonstrated that a simplified physical simulation of CBCT artifacts is sufficient to train a model that generalizes well to real CBCT data, eliminating the need for difficult-to-acquire real paired data.

4. Experimental Results

The method was evaluated on 75 prostate cancer patients (66 training, 8 testing).

A. Performance on Pseudo-CBCT (Efficiency & Quality)

Compression Analysis: Models with higher compression factors (smaller latent spaces) trained significantly faster.
- CLDM (128) (compression factor 4) achieved the best balance: 6 minutes/epoch for training and 30 seconds/slice for generation.
- This was **<1/8th** the training time of the non-compressed baseline (Palette) while maintaining high structural similarity (SSIM > 0.94).

B. Performance on Real CBCT (Quality & Structure)

The proposed method (CLDM Pseudo) was compared against:

GAN-based (Real/Pseudo): Trained on real or pseudo pairs.
CLDM (Real): Trained on real CT-CBCT pairs.

Key Findings:

Structural Preservation: The proposed method preserved anatomical structures almost perfectly.
- Structural Changes: The number of pixels showing structural changes (overcorrection) was < 1/1000th of the conventional CLDM (Real) method (3 pixels vs. 3,919 pixels in the test set).
- Qualitative: Unlike GAN and CLDM (Real) methods, the proposed method did not create hallucinated structures (e.g., turning soft tissue into gas) or distort organ shapes.
CT Value Accuracy:
- The correlation coefficient between the CT value distribution of the generated images and the reference CT was 0.916, comparable to conventional methods (0.908–0.934).
- Mean and standard deviation of CT values closely matched the reference CT.
Ablation Study: Removing any step from the pseudo-CBCT creation process (warping, contrast adjustment, or specific masks) resulted in lower correlation coefficients and persistent artifacts, confirming the necessity of the full simulation pipeline.

5. Significance

Clinical Impact: This method enables the direct calculation of radiation dose from CBCT images during treatment by improving image quality without compromising the anatomical integrity required for accurate registration.
Methodological Advancement: It proves that self-supervised learning with simulated data can outperform supervised learning with real, misaligned data in scenarios involving significant anatomical deformation.
Efficiency: By moving diffusion models to latent space, the study makes high-fidelity medical image enhancement computationally viable for clinical deployment, offering a faster alternative to standard diffusion models while maintaining superior stability compared to GANs.