MRI-guided adaptive radiotherapy (MRgART) is an advanced form of radiation treatment where a patient's anatomy is imaged with MRI immediately before each daily treatment session. Because the prostate, bladder, and rectum shift in position from day to day depending on bladder filling and bowel gas, this daily imaging allows the treatment plan to be adjusted to precisely match the patient's anatomy on that particular day, improving tumor coverage and reducing radiation exposure to surrounding organs.
The core technical challenge is that MRI does not provide electron density information, which is required to calculate exactly how radiation will travel through different tissues. CT scans do provide this information in the form of Hounsfield units, but acquiring a CT scan every treatment day is impractical due to radiation exposure, cost, and workflow complexity.
Current MRgART workflows solve this problem by deformably registering the original simulation CT scan to the daily MRI using deformable image registration (DIR), which mathematically warps the CT image to match the shape of the MRI. However, this registration is imperfect, particularly where gas pockets and bowel contents shift between the two scans. Manual density overrides must then correct the remaining mismatches, a process that takes approximately four minutes per patient per fraction and introduces inconsistency and user-dependent error.
An alternative approach is to generate a synthetic CT (sCT) directly from the daily MRI using deep learning, eliminating the need for registration and manual corrections entirely. This study investigated whether a generative adversarial network trained on low-field 0.35T MRI could produce synthetic CT images with sufficient accuracy for dose calculation in prostate stereotactic body radiotherapy.
Most synthetic CT research has focused on high-field 1.5T or 3T MRI scanners. The ViewRay MRIdian system used in this study operates at 0.35T, a substantially lower field strength chosen to allow integration with a linear accelerator without magnetic interference. Low-field MRI presents specific image quality challenges that make synthetic CT generation harder than at higher field strengths.
Lower field strength produces a lower signal-to-noise ratio, meaning images are inherently noisier. The system is also constrained to a TrueFISP imaging sequence (true fast imaging with steady-state precession) optimized for speed at low field, which does not provide the tissue-contrast diversity that high-field scanners can deliver through specialized pulse sequences such as Dixon or zero echo time imaging.
The fundamental difficulty is that both air and dense cortical bone appear as signal voids on MRI, making them visually indistinguishable even though they have dramatically different radiation attenuation properties. A deep learning model must learn to infer from surrounding anatomy which signal void represents air and which represents bone -- a challenge that MRI physics alone cannot resolve.
Prior work using 2D single-plane (typically axial) approaches for synthetic CT generation from low-field MRI reported mean absolute errors in CT numbers of 54 to 81 HU for whole pelvis regions and showed visible discontinuity artifacts between slices. The present study proposed a multi-planar approach combining axial, sagittal, and coronal MRI planes to reduce these discontinuities and improve overall accuracy.
Fifty-seven male patients who received MRI-guided pelvic radiation therapy at Brigham and Women's Hospital between 2019 and 2022 were enrolled. Each patient had both a CT simulation scan and a 0.35T MRI acquired on the same day in treatment position. Forty-five cases covering a range of pelvic treatment sites including prostate, pelvic nodes, and perirectal areas were used for training, and the remaining twelve prostate stereotactic body radiotherapy (SBRT) cases were reserved for validation and testing.
The deep learning model uses a conditional generative adversarial network (cGAN) based on the pix2pix architecture. A GAN consists of two competing neural networks: a generator that learns to produce realistic synthetic CT images from MRI input, and a discriminator that learns to distinguish generated synthetic CT from real CT. This adversarial training forces the generator to produce outputs that are not just numerically similar to real CT but that also have visually realistic texture and structure.
The generator was built on a U-Net architecture with skip connections that preserve fine spatial detail from the MRI input, while the discriminator used a PatchGAN structure that evaluates image realism in small local patches rather than globally. The combined loss function included three components: the adversarial loss driving realism, a pixel-level mean absolute error loss driving numerical CT number accuracy, and a mutual information loss added to handle misalignment uncertainty between the CT and MRI images acquired on the same day but not in identical patient positions.
All images were preprocessed to a standardized 256 x 256 x 256 volume at 1.5 x 1.5 x 1 mm resolution. MRI intensities were corrected for field non-uniformity using the N4itk bias field correction algorithm and normalized using a z-score method. CT intensities were normalized to a -1 to 1 range. The model was trained for 100 epochs using the Adam optimizer with different learning rates for the generator (0.0002) and discriminator (0.00005).
The key innovation in this study is the multi-planar training and inference strategy. Rather than training the network on only axial slices as most prior work has done, the network was trained simultaneously on 2D slices extracted from all three orthogonal orientations: axial, sagittal, and coronal. In each training iteration, one randomly selected slice from each orientation was used as a batch, exposing the generator to the 3D anatomy from multiple perspectives during training.
During inference (generating synthetic CT for a new patient), the trained generator was applied separately to each of the three orthogonal MRI slice sets, producing three complete synthetic CT volumes: sCTax from axial slices, sCTsag from sagittal slices, and sCTcor from coronal slices. Each of these three volumes has higher quality in the plane from which it was generated but shows discontinuity artifacts in the perpendicular planes.
The final synthetic CT volume (sCTave) was obtained by averaging the three volumes voxel-by-voxel. This averaging step is the mechanism by which the multi-planar approach reduces discontinuities: artifacts that appear in one plane's output but not the others are averaged out, while true anatomical signal that is consistent across all three orientations is preserved and reinforced.
The entire synthetic CT generation process required approximately 12 seconds of computation on a GPU, compared to approximately 4 minutes for the current clinical workflow of deformable image registration plus manual density overrides. This speed advantage is a practical prerequisite for integration into the time-constrained online adaptive radiotherapy workflow, where treatment must proceed within a limited on-table window.
For the 12 validation cases, the averaged multi-planar synthetic CT (sCTave) achieved a mean absolute error of 30.1 plus or minus 4.2 HU across the whole pelvis region, a peak signal-to-noise ratio of 35.2 plus or minus 1.7, and a structural similarity index of 0.9758 plus or minus 0.0035. These metrics were all superior to those from the single-plane axial (32.8 HU), sagittal (33.6 HU), and coronal (33.2 HU) approaches, confirming the benefit of multi-planar averaging.
Accuracy varied substantially by tissue type. Soft tissue was predicted most accurately with an MAE of only 19.6 HU, reflecting the relatively good MRI contrast between different soft tissues. Bone prediction was harder at an MAE of 158.5 HU, consistent with the known challenge of inferring bone density from MRI signal voids. The largest errors appeared in air segments (MAE of 396.5 HU), primarily because gas pocket positions in the bowel and rectum shifted between the CT simulation and MRI acquisition times even on the same day.
Compared to a prior study using a 2D cGAN on the same 0.35T MRI system, the multi-planar method reduced whole pelvis MAE from 54.3 HU to 30.1 HU, soft tissue MAE from 40.4 HU to 19.6 HU, and bone MAE from 224.4 HU to 158.5 HU -- approximately 45, 51, and 29 percent improvements respectively. Results were also comparable to or better than published synthetic CT methods using higher-field 1.5T and 3T MRI for pelvic applications.
Visual inspection confirmed that the sCTave images had smoother edges in all planes compared to the single-orientation outputs, with reduced slice discontinuity artifacts. The synthetic CT images were visually comparable to the deformed CT images used as reference, with the most notable differences occurring adjacent to gas pockets where bowel positions had shifted between acquisitions.
Clinical treatment plans originally calculated on deformed CT images were recalculated on the synthetic CT volumes using the same beam configurations and Monte Carlo dose calculation algorithm, which accounts for the effect of the 0.35T magnetic field on radiation transport. This allowed direct comparison of the dose distributions predicted by synthetic CT versus the reference clinical plans.
The mean dosimetric difference was less than 1 percent for all evaluated planning target volume and organ-at-risk metrics. Mean dose differences were within 0.25 Gy for PTV dose metrics and within 0.35 Gy for OAR metrics (rectum, bladder, and urethra), representing 0.6 percent and 0.9 percent of the 36.25 Gy prostate SBRT prescription dose respectively.
Three-dimensional gamma analysis, which simultaneously evaluates spatial and dose agreement between two dose distributions, achieved pass rates of 99 percent for 1%/1 mm criteria and 99.9 percent for 2%/2 mm criteria across all validation cases. These pass rates are consistent with clinical acceptance standards and confirm that synthetic CT-based dose calculations are reliable enough for clinical use.
The worst-performing case involved a large gas pocket in the rectum adjacent to the prostate. This case produced a rectal D2% deviation of 0.9 Gy (2.3%), the only case exceeding 0.7 Gy deviation. The difficulty in this case arose from the electron return effect, a phenomenon where the MRIdian's magnetic field causes electrons scattered at air-tissue interfaces to travel back into the tissue, amplifying dose near gas cavities. Accurately modeling this effect requires precise knowledge of gas pocket geometry, which cannot be reliably inferred from MRI signal voids alone.
This study demonstrated that multi-planar cGAN-generated synthetic CT images from 0.35T MRI can achieve dosimetric accuracy within 1 percent of conventional deformed CT methods for prostate SBRT, while being generated in 12 seconds instead of 4 minutes. This speed improvement is clinically significant for online adaptive radiotherapy, where the total time on-table is constrained and any step that slows the workflow reduces its feasibility.
By replacing deformable image registration and manual density overrides with a single automated deep learning inference step, the synthetic CT approach would eliminate a major source of user-dependent variability in the adaptive radiotherapy process. Consistency in density assignment is particularly important for accurate dose calculation near tissue-air interfaces where the magnetic field effects are strongest.
The main acknowledged limitation is the small validation cohort of 12 cases from a single institution, and the exclusion of patients with hip implants or excessive imaging artifacts, who represent a clinically common population that any deployed system must handle. The image resolution used also limits the field of view to 38.4 x 38.4 cm, potentially cutting off anatomy for larger patients.
Future work will validate the method in a larger patient cohort, extend it to other body sites where MRgART is used, and develop a quality assurance process to automatically flag outlier synthetic CT cases for physician review before treatment proceeds. Integration of such a QA step would be essential for clinical deployment, ensuring that the rare cases where the deep learning prediction fails are caught before they affect treatment delivery.