MRI scan transforming into a CT image with AI neural connections

MRI-Only Prostate IMRT: Can AI Deliver Accurate Synthetic CT Images?

Reese Marlowe monogram
Reese Marlowe in Health & Wellness October 2026 • 4 min read.

"A deep learning U-net model shows promise in generating synthetic CT images from MRI for prostate cancer radiotherapy, potentially streamlining treatment planning and reducing radiation exposure."


Magnetic Resonance Imaging (MRI) has become a cornerstone in radiotherapy, offering superior soft tissue contrast compared to Computed Tomography (CT). This makes MRI ideal for delineating organs. Traditionally, MRI serves as a secondary tool, with images rigidly aligned to simulation CT scans for tumor and organ definition.

However, a fully MRI-guided radiotherapy workflow holds the potential to eliminate image registration uncertainties, reduce radiation exposure, and enhance patient comfort. The key lies in deriving CT-equivalent information, such as electron density distribution, directly from MR images for accurate radiation dose calculation. This data is commonly referred to as pseudo-CT or synthetic-CT (S-CT) images.

Researchers have been exploring various methods to generate S-CT images, including statistical modeling, multi-atlas deformable image registration, random forest-based approaches, and deep learning techniques. This article delves into a recent study investigating the feasibility of training a deep convolutional neural network (ConvNet) from scratch to generate S-CT images from conventional T2-weighted MR images, specifically for prostate cancer treatment planning.

AI Search Multiple angles on this topic

MRI-Guided Radiotherapy for Prostate Cancer

Magnetic resonance-guided radiotherapy (MRgRT) became possible with the clinical introduction of an integrated MRI and linear accelerator, or MR-LINAC. A narrative review describes it as a novel technique for prostate cancer treatment. The review says it promises to improve clinical outcomes and reduce toxicity, but the source does not provide statistics quantifying those effects.

Standard Image Guidance and Its Limits

External beam radiotherapy is described as an optimal treatment choice for men with localized prostate cancer and as providing long-term disease control in most patients. Image-guided prostate radiotherapy is standard of care. However, current techniques can involve invasive procedures and imaging with poor soft-tissue resolution, which can limit accuracy.

U-Net: A Deep Learning Solution for Synthetic CT Generation

MRI scan transforming into a CT image with AI neural connections

The study, recently published, details the development and evaluation of a U-net based deep learning model for generating synthetic CT images from MRI data in prostate cancer patients. Here’s a breakdown of their approach:

Researchers used paired CT and T2-weighted MR images from 51 prostate cancer patients. They divided the data into a training set (36 patients) and a testing set (15 patients). The training data was augmented using artificial deformations to improve the model's robustness.

  • U-net Architecture: A 2-dimensional U-net, featuring 23 convolutional layers and 25.29 million trainable parameters, was employed. The U-net learns a nonlinear function to map MR image slices to corresponding S-CT slices.
  • Evaluation Metrics: The mean absolute error (MAE) of Hounsfield Units (HU) between true CT and S-CT images was used to assess HU estimation accuracy.
  • IMRT Planning: Intensity-modulated radiation therapy (IMRT) plans, prescribing a dose of 79.2Gy to the planning target volume (PTV), were created using the true CT images. These plans were then recalculated using the S-CT images.
  • Dosimetric Comparison: The dose matrices from the true CT and S-CT-based plans were compared using gamma index analysis and absolute point dose discrepancy measurements.
AI Search Multiple angles on this topic

Recent Reviews and Two-Year Findings

A systematic review published April 1, 2026, highlights potential clinical benefits of MRI-guided linear accelerators for prostate cancer radiotherapy across a variety of clinical indications. Separately, a report published March 17, 2025, says MRI-guided stereotactic body radiotherapy (SBRT) was associated with reduced long-term gastrointestinal and genitourinary toxicity compared with CT-guided SBRT after two years. That report also links MRI guidance to less deterioration in bowel and sexual function.

MRI Versus CT Guidance

A comparison of prostate radiotherapy guidance methods reports that MRI guidance using smaller margins resulted in lower rates of side effects than standard CT guidance. Updated results were reported at two years after treatment. The MIRAGE trial compared MRI-guided daily adaptive prostate SBRT with standard CT-guided SBRT and used MRI image guidance to achieve aggressive planning target volume margin reduction.

The U-net model was trained from scratch using a GP100-GPU over 58.67 hours. Generating a new S-CT volume image took only 3.84 to 7.65 seconds. The mean MAE within the body was 29.96 ± 4.87 HU. Gamma pass rates (1%/1mm and 2%/2mm) exceeded 98.03% and 99.36%, respectively. DVH parameter discrepancies were less than 0.87%, and the maximum point dose discrepancy within the PTV was less than 1.01%.

The Future of MRI-Only Radiotherapy

The study demonstrates the potential of deep learning, specifically the U-net architecture, to generate accurate S-CT images from conventional MR images in seconds. This opens doors for streamlined MRI-only radiotherapy workflows in prostate cancer treatment.

AI Search Multiple angles on this topic

Emerging MRI Technologies and AI

A review published May 22, 2026, describes new advancements in prostate MRI technology. It points to novel imaging sequences and the application of machine learning and artificial intelligence across the imaging pipeline. These developments suggest further possibilities for the field, according to the review.

Acquisition Time and Image Quality

Efforts to shorten MRI acquisition time have led to short-MRI protocols and quality scores such as PI-quality. Radiomics has also gained importance in prostate imaging. The source notes that multiparametric MRI’s role in prostate cancer diagnosis is expanding.

While the U-net model shows promising results, the researchers acknowledge areas for further improvement. These include incorporating more anatomical information into the training data, exploring advanced network architectures like 3D U-nets, and utilizing adversarial training techniques to generate more realistic images.

Despite these limitations, this research represents a significant step toward realizing the full potential of MRI-only radiotherapy, offering a faster, more accurate, and less invasive approach to prostate cancer treatment planning.

Synthetic CT accuracy beside other approaches

  • 20With 20 training patients, a U-net model reported whole-body mean absolute error of 29.68 ± 4.41 HU; the same measure was 35–66 HU across cited pelvic-region studies.[1] [2]
  • 0.3%In a multicenter, multivendor validation, mean dose deviations across all DVH metrics were below 0.3%.[3]

What the comparisons say about synthetic CT

The reported whole-body error falls within the range cited for pelvic-region work, but those figures come from different studies and do not establish that their models or evaluations were equivalent.[1] [2]

Dose comparisons suggest that synthetic CT can produce dose estimates close to CT-based estimates in some evaluations. That does not establish reliability for every patient or MRI acquisition.[3] [4]

How synthetic CT compares across prostate workflows

These workflow figures show that synthetic-CT generation has been used beyond image-accuracy testing, while also documenting cases where reconstruction did not succeed; they are not a guarantee for a different model or setting.[5]

What remains uncertain in MRI-only planning

  • How well does a model retain dose-calculation reliability when a patient's anatomy differs from its training cohort, or when MRI artifacts, sequences, or devices differ?[4] [6]
  • How can contour-related registration uncertainty be addressed when MRI-defined prostate contours may not align with CT fiducial positions?[7]
  • What range of Hounsfield-unit uncertainty can be tolerated without causing dose-distribution errors, and how should automated methods identify that range?[8]
  • How often do motion or phase-wrap artifacts affecting the outer body contour, or patients larger than atlas cases, lead to inaccurate tissue-density estimates?[6]

Sources

  1. Deep learning‐based synthetic CT generation for MR‐only radiotherapy of prostate cancer patients with 0.35T MRI linear accelerator (pmc.ncbi.nlm.nih.gov)
  2. A deep learning model to generate synthetic CT for prostate MR-only radiotherapy dose planning: a multicenter study (pmc.ncbi.nlm.nih.gov)
  3. Cone beam CT for QA of synthetic CT in MRI only for prostate patients (pmc.ncbi.nlm.nih.gov)
  4. Quality assurance for MRI-only radiation therapy: A voxel-wise population-based methodology for image and dose assessment of synthetic CT generation methods (www.frontiersin.org)
  5. Clinical experience and workflow challenges with magnetic resonance-only radiation therapy simulation and planning for prostate cancer (pmc.ncbi.nlm.nih.gov)
  6. Cone beam computed tomography for dose calculation quality assurance for magnetic resonance-only radiotherapy (pmc.ncbi.nlm.nih.gov)
  7. A Multi-center Prospective Study for Implementation of an MRI-Only Prostate Treatment Planning Workflow (pmc.ncbi.nlm.nih.gov)
  8. Determination of acceptable Hounsfield units uncertainties via a sensitivity analysis for an accurate dose calculation in the context of prostate MRI-only radiotherapy (pmc.ncbi.nlm.nih.gov)

About this Article -

Written with AI assistance from published research, and reviewed by the Mystum team. See our About page for more information.

Everything You Need To Know

1

What exactly is the U-net, and how does it relate to prostate cancer treatment?

The U-net is a deep learning model used to generate synthetic CT (S-CT) images from MRI data, specifically for prostate cancer treatment. It is a type of convolutional neural network (ConvNet) that excels at image segmentation and reconstruction tasks. In this context, the U-net learns to map MR image slices to corresponding S-CT slices. This is significant because it allows for the creation of CT-equivalent information from MRI images, which is crucial for accurate radiation dose calculation in Intensity-modulated radiation therapy (IMRT).

2

What are Synthetic CT (S-CT) images, and why are they important?

Synthetic CT (S-CT) images are CT-like images generated from MRI data. They are crucial because they provide information about electron density distribution, which is essential for accurate radiation dose calculation in radiotherapy. Traditionally, MRI is used alongside CT, but S-CT images allow for a fully MRI-guided radiotherapy workflow, potentially eliminating image registration uncertainties and reducing radiation exposure for patients. The process involves a deep learning model, like the U-net, which is trained to convert MRI data into S-CT images.

3

Why is MRI used, and how does the U-net help in its application?

MRI (Magnetic Resonance Imaging) offers superior soft tissue contrast compared to CT (Computed Tomography), making it ideal for delineating organs. However, MRI alone cannot provide the electron density information needed for radiation dose calculation. In prostate cancer radiotherapy, MRI is increasingly used for treatment planning. The U-net model enables the use of MRI by generating S-CT images which allows for accurate radiation dose calculation by the IMRT planning system. Thus, it streamlines the process and reduces radiation exposure to the patient.

4

How does the U-net fit into the process of Intensity-modulated radiation therapy (IMRT)?

Intensity-modulated radiation therapy (IMRT) is a sophisticated radiotherapy technique that uses CT scans to deliver precise radiation doses to the tumor while sparing surrounding healthy tissues. The U-net model plays a crucial role in IMRT by generating synthetic CT images from MRI data. This allows for creating IMRT plans directly from MRI scans, potentially eliminating the need for CT scans in the treatment planning process. These IMRT plans are then recalculated using the S-CT images to ensure that the radiation dose is delivered accurately.

5

Can you explain the complete process of using the U-net for prostate cancer treatment?

The process involves several steps. First, a U-net deep learning model is trained using paired CT and T2-weighted MR images. This training process involves using a training set of images and optimizing the model's ability to map MR images to S-CT images. Once trained, the U-net can generate S-CT images from new MRI data within seconds. These generated S-CT images are then used for IMRT planning and dose calculation. The results are then evaluated using metrics such as mean absolute error (MAE) and gamma pass rates to ensure accuracy and precision in radiation delivery. The model uses 23 convolutional layers and 25.29 million trainable parameters.

Newsletter Subscribe

Subscribe to get the latest articles and insights directly in your inbox.