External beam radiotherapy (EBRT) for prostate cancer has traditionally required two separate imaging steps: a CT scan to calculate radiation dose (because CT accurately measures tissue density) and an MRI for superior soft-tissue delineation of the prostate and nearby organs. Patients must attend both sessions, and their images must be registered together -- a process prone to small but consequential misalignments.
CT simulation exposes patients to additional ionizing radiation and adds time and cost to an already complex treatment pathway. For prostate radiotherapy specifically, CT provides poor contrast between the prostate and surrounding tissues, making MRI indispensable for target delineation. Yet CT cannot simply be discarded because its Hounsfield unit (HU) values are used to calculate how much radiation reaches each tissue layer.
The solution is a synthetic CT (sCT): a deep learning model that converts MRI images into CT-like images with realistic tissue density values, allowing dose calculation without an actual CT scan. This enables a fully MRI-only radiotherapy planning (MROP) workflow in which a single MRI session replaces both imaging modalities, eliminating registration errors, reducing radiation exposure, and streamlining scheduling.
This paper describes the complete development and clinical implementation of an MROP workflow at a single institution, covering the deep learning sCT generation algorithm, automated fiducial marker detection, prospective validation in patients, and outcomes from over 500 patients treated using this approach since its clinical launch in June 2023.
The sCT generation model used a multi-channel CycleGAN architecture -- a type of generative adversarial network that learns to translate images between domains (in this case, MRI to CT) without requiring perfectly aligned image pairs. The model used two MRI channels as input: T1-weighted fat images (hyperintense in fat tissue) and T1-weighted in-phase (IP) images (hyperintense in both fat and bone), both derived from a T1 Dixon sequence.
Using both fat and IP channels was deliberate: fat images delineate soft tissue boundaries, while in-phase images highlight bone. Because accurate bone representation is critical for dose calculation, a bony structure constraint was added to the CycleGAN loss function, penalizing errors in bone density reproduction specifically. This modification resulted in the best bone Dice similarity coefficient (DSC) of 0.88 and the lowest whole-body mean absolute error (MAE) of 50.7 HU in retrospective evaluation.
The workflow required a 59.5-minute MRI simulation session, slightly longer than MRI alone but saving the 32-minute CT simulation entirely. The MRI protocol included T2-weighted images for contouring, T1 Dixon images for sCT generation, and a quantitative susceptibility mapping (QSM) sequence using multi-echo T2*-weighted images for fiducial marker detection.
Post-processing scripts using application programming interfaces (APIs) were developed to embed fiducial marker locations into the sCT at high HU values, insert immobilization frame geometry, and apply automatic bladder density override -- ensuring the bladder always received a water-equivalent density value regardless of variations in bladder MRI signal intensity. These automated steps addressed practical failure modes identified during clinical implementation.
Prostate radiotherapy uses implanted fiducial markers (FMs) -- small gold seeds placed in the prostate -- as reference points for precise daily patient alignment during treatment. These markers are clearly visible on CT but can be difficult to distinguish from calcifications or air pockets on MRI, as all three appear as signal voids on standard sequences.
Quantitative susceptibility mapping (QSM), derived from multi-echo T2*-weighted MRI, was used to solve this problem. QSM measures the magnetic susceptibility of each tissue, and the gold FMs have a distinct susceptibility of -36 plus or minus 4 ppm, compared to air (+0.024 ppm) and calcification (-21 plus or minus 3 ppm). This physical difference allows the QSM algorithm to automatically identify markers while rejecting false-positive candidates from normal anatomy.
In a validation cohort of 22 patients with 66 fiducial markers, QSM achieved an initial detection rate of 89.4%, compared to 83.3% for manual identification by an experienced physicist. When the physicist reviewed images with access to QSM-generated marker locations, the combined detection rate increased to 95.5% (63 of 66 markers) -- the three undetected markers were at the prostate edge. Critically, average detection time per patient was reduced from approximately 10 minutes to 2 minutes.
A quality assurance (QA) workflow was implemented in which automatically detected FM locations are cross-validated against signal voids on T1 Dixon water images and then transferred to the T2-weighted planning images for final verification. Manual co-registration between sequences was required in only 5 of approximately 500 patients, typically due to organ movement caused by bladder filling changes during the 20-minute MRI examination.
Following retrospective validation, 12 consecutive patients were enrolled in prospective evaluation of the MROP workflow. Ten completed treatment using MROP; one was excluded due to inadequate bladder and rectal filling at simulation, and one had sCT artifacts in the prostate region and the treating physician elected to use conventional CT. These cases informed specific workflow improvements implemented afterward.
Across the 10 prospective patients, the mean error (ME) between sCT and planning CT was -0.25 plus or minus 5.90 HU for the entire body, indicating minimal systematic bias. Bony structures showed the largest deviations (ME of 55.2 plus or minus 24.3 HU), consistent with the known difficulty of reproducing cortical bone CT numbers precisely from soft-tissue MRI sequences.
For radiation dose accuracy, all planning target volumes (PTVs) showed average dose differences below 1% between MROP-based and CT-based treatment plans. This held across prostate-only plans, plans with simultaneous boost targets, and plans including regional lymph node volumes -- confirming the sCT is accurate enough for the full range of clinical prostate treatment scenarios. Dose differences to organs at risk (bladder, rectum, bowel) were within acceptable clinical ranges.
Two artifact types were identified during prospective use and subsequently corrected: large-diameter fiducial markers created contrast-enhanced spots inside the prostate on sCT, and excessive rectal gas caused susceptibility artifacts extending into the prostate region. Model refinements incorporating bladder segmentation as guidance were implemented, and since these improvements only 8 of approximately 350 patients required any manual density override before treatment planning.
After successfully completing the prospective 10-patient evaluation, the MROP workflow received departmental approval and CT simulation was permanently removed from the prostate treatment pathway. This administrative decision was supported by outcome data and represented a fundamental change to clinical practice that required buy-in from physicians, physicists, dosimetrists, and radiation therapists.
Removing CT simulation simplified scheduling: patients no longer needed to find a day when both CT and MRI suites were available back-to-back. This change reduced average patient wait time from physician consultation to simulation by 7.1 days in the first 6 months of clinical operation compared to the preceding 6 months. Total simulation time was also reduced by an average of 21 minutes per patient, as the MROP MRI simulation (59.5 min) replaced two separate sessions (CT: 32 min plus MRI: 49 min).
The workflow was also extended to the MR-Linac (MRL) -- a hybrid radiation delivery system that combines MRI with treatment delivery for real-time tumor tracking. The MRL workflow differs from the C-arm Linac approach: T2-weighted images are used for daily treatment guidance, and bone density is assigned from the sCT on a per-contour basis rather than using the sCT directly for planning. Five patients were treated on the MRL using MROP before the workflow was formally approved for clinical use on that platform.
Since clinical go-live in June 2023, over 500 patients have been treated using MROP on both C-arm Linac and MRL systems. Approximately 7% of patients are ineligible for MROP, primarily due to metal implants (5%), patient diameter exceeding coil coverage (1%), or severe bowel gas artifacts (1%). An additional 7% require supplementary manual steps such as density overrides or additional CT verification for fiducial marker confirmation.
The MROP workflow succeeded because implementation followed a structured progression: workflow design, technical development, retrospective validation, prospective pilot, clinical go-live, and continuous quality improvement. This phased approach allowed issues to be identified and corrected before large numbers of patients were affected, and it mirrors the responsible translation of AI-based tools into clinical practice more broadly.
An important advantage of the institution's custom deep learning approach over commercial sCT software was the extended field of view: 400 mm superior-to-inferior coverage versus 240 mm for the standard vendor solution, and a 550 x 550 mm transverse FOV versus 450 x 450 mm. This extended coverage is essential for patients with both prostate and nodal disease, where treatment planning fields extend well beyond the pelvis.
Known limitations include long MRI scan times that may cause organ motion between sequences, challenges in patients with large body habitus or metal implants, and variability in fiducial marker type across urological practices -- a problem that emerged when some urologists adopted a newer marker type less conspicuous on current MRI sequences. Each of these limitations represents a target for future technical development, including accelerated MRI acquisition protocols and sequence updates for new marker types.
The institution is now working to extend the sCT generation framework beyond the prostate to other body sites. The experience accumulated with prostate MROP -- including patient selection criteria, organ motion management, fiducial detection strategy, and workflow quality assurance -- provides a reusable foundation for implementing MRI-only planning in thoracic, abdominal, and other pelvic treatment sites.