Radiation therapy for cancers like prostate cancer requires extremely precise treatment planning to direct high doses of radiation at tumors while sparing healthy tissue. This planning relies on medical imaging to map the body accurately.
MRI (Magnetic Resonance Imaging) is excellent at showing soft tissue detail, which makes it ideal for distinguishing tumors from surrounding organs -- especially in the pelvis and brain. However, MRI cannot directly measure tissue density, which is essential for calculating how radiation will interact with the body.
CT (Computed Tomography) provides the precise density measurements needed for radiation dose calculations but involves ionizing radiation exposure and provides less soft tissue contrast than MRI. Patients currently must undergo both scans, which adds cost, time, and additional radiation exposure.
Generating a synthetic CT (sCT) directly from an MRI scan would allow clinicians to use only MRI for treatment planning -- eliminating the CT scan, reducing radiation dose to patients, and removing the errors that occur when trying to align two separately acquired scans.
Previous attempts at MRI-to-CT translation have used Generative Adversarial Networks (GANs) -- a type of AI where two neural networks compete against each other to produce realistic-looking images. GAN-based methods can generate visually plausible synthetic CTs quickly, but they suffer from unstable training and may produce images that look realistic but contain subtle inaccuracies.
Diffusion models are a newer class of AI image generators inspired by physics. They work by first gradually adding random noise to an image until it becomes pure static, then learning to reverse that process to reconstruct the original. This approach produces more accurate and diverse outputs than GANs without the instability problems.
However, existing diffusion models used for medical imaging had a major drawback: they were very slow. Generating a 3D volume slice by slice using 2D diffusion models lost important spatial relationships between slices, while a fully 3D approach was computationally expensive and impractically slow for clinical use.
The challenge was to build a 3D diffusion model that could generate high-quality synthetic CTs quickly enough to be used in real clinical workflows.
The new model, called MC-IDDPM (MRI-to-CT Improved Denoising Diffusion Probabilistic Model), uses a two-stage process. In the forward pass, a real CT scan is gradually corrupted by adding random Gaussian noise over many small steps until it becomes pure static -- essentially turning structured information into random noise.
In the reverse pass, a neural network learns to undo this process, step by step, guided by the patient's MRI scan. The network predicts what the noise-free CT would look like, conditioned on the anatomy visible in the MRI. Each denoising step produces a slightly cleaner image until a realistic synthetic CT is generated.
A key innovation over older diffusion models is that MC-IDDPM predicts both the mean and variance of the noise at each step -- not just the average. This allows the model to take much larger steps during generation, reducing the number of required iterations by about 20 times compared to standard diffusion models.
To add another layer of reliability, the model uses Monte Carlo generation -- running the process 5 times per patient and averaging the results. This reduces the randomness inherent in the diffusion process and produces a more stable, reliable final image.
The backbone of MC-IDDPM is a specialized neural network called Swin-Vnet, which combines two powerful architectural ideas: Shifted-Window (Swin) Transformer attention and the V-net 3D architecture originally designed for medical image segmentation.
Transformer attention mechanisms allow the network to capture long-range relationships within an image -- understanding how structures in one part of the scan relate to structures far away. This is especially important in medical imaging, where the spatial relationship between organs provides critical context.
The shifted-window approach makes transformer attention computationally practical for 3D medical volumes. Rather than computing attention across the entire image at once (which would be prohibitively expensive), it processes the image in overlapping windows that shift between layers, allowing global context while keeping computation manageable.
The model was trained on patch-based inputs -- small 3D sub-volumes of the full scan -- and used a sliding window approach during inference, enabling it to handle full clinical-resolution scans despite memory limitations of available hardware.
The model was tested on two clinical datasets: 36 patients with brain imaging and 28 patients with prostate imaging. Performance was measured using four standard metrics: Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Normalized Cross Correlation (NCC).
For brain imaging, MC-IDDPM achieved an MAE of 48.8 HU (Hounsfield Units -- the standard measure of CT image intensity accuracy), SSIM of 0.947, and NCC of 0.976 -- outperforming all six competing methods with statistically significant improvements (p < 0.05) on most metrics.
For prostate imaging, the model achieved an MAE of 55.1 HU, PSNR of 28.7 dB, SSIM of 0.878, and NCC of 0.940 -- again outperforming GAN-based methods by large margins. Compared to competing diffusion models, MC-IDDPM showed the most consistent improvements, especially in the MAE metric that directly reflects CT accuracy.
These results indicate that MC-IDDPM produces synthetic CTs that are substantially closer to real CT scans in both absolute intensity values and overall structural appearance, compared to previous state-of-the-art methods.
Image quality metrics alone do not tell us whether a synthetic CT is accurate enough for radiation treatment planning. The critical test is whether a treatment plan calculated using the sCT would deliver the same radiation dose as a plan calculated on a real CT scan.
The research team performed a detailed dosimetric analysis for a brain cancer patient with four separate Planning Target Volumes (PTVs) -- distinct tumor regions each requiring precise radiation doses. They compared dose distributions calculated using the synthetic CT versus the real planning CT.
The results showed that dose coverage differences were within 0.34% across all planning target volumes for both D95% and D99% metrics (the dose received by 95% and 99% of the target volume). This level of agreement is well within the clinical tolerance limits used in radiation therapy.
This finding is highly significant because it validates that MC-IDDPM's synthetic CTs are not just visually similar to real CTs -- they are accurate enough for the precise dose calculations that radiation therapy planning requires. The sCT could potentially replace the CT scan entirely in the treatment workflow.
The main limitation of MC-IDDPM is its generation speed. Using 50 sampling steps and 5 Monte Carlo runs, the model takes approximately 760 seconds (about 13 minutes) to generate a synthetic CT volume. GAN-based methods can produce the same resolution in just 22 seconds.
This speed limitation arises from two sources: the need for multiple denoising iterations to achieve high quality, and the need for multiple Monte Carlo runs to average out random variation. These are fundamental properties of the diffusion approach rather than hardware constraints.
Future work will focus on efficiency improvements through algorithmic advances such as exponential forward processes that require fewer timesteps, or pre-trained networks that can accelerate image synthesis. Deterministic diffusion processes that eliminate the need for multiple Monte Carlo runs are also being explored.
The researchers also plan to extend MC-IDDPM to other medical imaging applications beyond MRI-to-CT translation, including ultrasound and cone beam CT, which have not yet been explored with 3D diffusion-based approaches.