Radiation therapy planning (RTP) requires precise knowledge of tissue density to calculate how much radiation reaches a tumor and how much is absorbed by surrounding healthy tissue. CT (computed tomography) provides this through measurements called Hounsfield Units (HU), which represent the electron density of each tissue type. This makes CT the gold standard for dose calculation in most countries.
The problem is that CT exposes patients to ionizing radiation, which is particularly concerning when multiple scans are needed to track treatment response. CT also provides poor contrast between soft tissues, making it difficult to precisely outline tumors in regions like the prostate, uterus, or pelvic structures where accurate boundary definition is critical.
MRI (magnetic resonance imaging) solves the soft tissue contrast problem -- it can clearly distinguish tumor from adjacent normal tissue and is far superior to CT for target delineation and for defining organs at risk (OARs) that must be spared from radiation. However, MRI does not directly measure tissue density and therefore cannot be used alone for dose calculations.
The solution being pursued is synthetic CT (sCT): using algorithms to generate a CT-like image from MRI data. If successful, this would enable MRI-only radiotherapy workflows -- eliminating the need for a separate CT scan, reducing radiation exposure, simplifying logistics, and eliminating errors from imperfect image registration between the two modalities.
The authors conducted a systematic literature review following PRISMA guidelines, searching major databases (PubMed, Scopus, Google Scholar, IEEE Xplore, ScienceDirect, arXiv) for studies published between 2013 and 2023. Starting from 162 identified papers, they reduced the set to 52 studies through three rounds of screening: first checking titles and abstracts, then full-text relevance, then final categorization.
The review focused exclusively on peer-reviewed journal articles and highly cited conference papers describing methods for generating sCT from MRI for radiotherapy applications. Book chapters, dissertations, and non-English articles were excluded. Studies were then grouped into five major categories based on their approach: sequence-based, segmentation-based, atlas-based, machine learning (ML)-based, and deep learning (DL)-based methods.
The most studied tumor sites in the included papers were brain, head and neck, prostate, pelvis, and abdomen -- regions where either MRI's soft tissue contrast is particularly valuable or where radiation therapy is commonly used. Prostate cancer was one of the most frequently studied sites, reflecting clinical demand for MRI-only planning in this anatomy.
Sequence-based methods exploit the fact that different MRI acquisition protocols (T1-weighted, T2-weighted, Dixon, ultrashort echo time) produce different signal intensities for different tissue types. By mathematically mapping these signal variations to known HU values, the method can estimate electron density without needing a reference CT scan. The main advantage is simplicity and no requirement for pre-existing CT data; the drawback is difficulty distinguishing bone from soft tissue.
Segmentation-based methods first divide the MRI image into distinct tissue classes (bone, fat, soft tissue, air) using techniques like fuzzy c-means clustering. Each class is then assigned a fixed or statistically derived HU value. This approach is conceptually straightforward but tends to produce step-function images with abrupt transitions between tissue types, which can introduce dosimetric errors at boundaries.
Atlas-based methods maintain a library of paired MRI-CT scans from prior patients. When a new patient's MRI arrives, the system finds the best-matching atlas scans through deformable image registration -- mathematically warping atlas images to align with the patient's anatomy. The corresponding CT HU values from matched atlas scans are then transferred to the patient. Multi-atlas approaches use several matched atlases and combine them, reducing errors from any single imperfect match.
All traditional methods have significant limitations: segmentation errors propagate into HU assignments, atlas-based methods struggle with anatomical variants not well represented in the library, and sequence-based methods require acquisition of multiple MRI sequences which increases scan time and patient discomfort. These limitations motivated the move toward AI-based approaches.
Generative adversarial networks (GANs) have become the dominant deep learning approach for sCT generation. A GAN consists of two competing neural networks: a generator that learns to produce realistic-looking CT images from MRI input, and a discriminator that tries to distinguish real CT scans from generated ones. Through this adversarial training, the generator improves until its outputs are indistinguishable from real CT scans. GAN-generated sCTs have consistently shown better visual realism and lower mean absolute error (MAE) compared to earlier methods.
CycleGAN is a variant that does not require perfectly paired MRI-CT image sets for training -- it learns the mapping between the two domains using unpaired images. This is practically valuable because acquiring precisely aligned paired scans is logistically challenging in clinical settings. U-Net architectures -- encoder-decoder networks with skip connections -- have achieved MAE values as low as 17.6 HU in soft tissue for brain applications and provide excellent spatial resolution in the generated images.
Residual neural networks (ResNets) use shortcut connections that allow gradient information to flow directly across multiple layers, enabling training of much deeper networks. For sCT generation, a 34-layer residual U-shaped network (RUN) demonstrated high accuracy across brain and prostate applications. Residual GAN models outperformed standard U-Net architectures in head and neck cancer sCT generation, achieving the lowest MAE and highest peak signal-to-noise ratio.
A newer approach uses diffusion probabilistic models -- a class of generative AI that progressively adds noise to CT images during training, then learns to reverse this process from MRI input. Applied to prostate and brain sCT generation, this approach showed promising results but is computationally expensive due to its long multi-step generation chain. Overall, deep learning methods generate higher-accuracy sCTs, are adaptable to different anatomies, and are fast enough for potential clinical deployment.
MRI-only radiotherapy is the most direct application, eliminating the separate CT scan from the treatment planning workflow. Current clinical focus is on prostate cancer (where MRI's soft tissue contrast makes it ideal for defining the target and nearby rectum and bladder), head and neck cancer, and brain tumors. Clinical validation studies have shown that sCT-based dose distributions are statistically equivalent to CT-based plans in these sites, with dose-volume histogram agreement within 0.30% across all structures.
PET attenuation correction is another key application. PET images measure metabolic activity but require correction for the way tissue absorbs the emitted radiation on its way out of the body. Traditionally this used CT data, which added radiation dose. Using sCT generated from the PET scanner's built-in MRI (in hybrid PET/MR systems) allows attenuation correction without additional CT, reducing dose and improving workflow in combined imaging.
Adaptive radiotherapy is perhaps the most transformative application: because anatomy changes during a treatment course (tumors shrink, patients lose weight, organs shift), the optimal treatment plan should be updated periodically. sCT enables replanning based on MRI scans acquired at each fraction of treatment, without repeatedly CT-scanning patients. This approach, already implemented on MRI-linear accelerator hybrid machines, allows real-time adaptation of dose delivery.
Commercially available products from major vendors including Philips (MRCAT for prostate, brain, and pelvis), Siemens Healthineers, GE Healthcare, Elekta (MR-Linac), and ViewRay (MRIdian) have integrated sCT generation into clinical systems. This commercial adoption signals that sCT technology is transitioning from research prototype to routine clinical tool, particularly in high-resource settings.
Each method category has distinct trade-offs. Sequence-based approaches integrate easily into existing clinical workflows and need no CT reference data, but produce lower accuracy and struggle with bone-soft tissue boundaries. Atlas-based methods are more accurate but require large, diverse training databases and fail when patient anatomy deviates significantly from available atlases -- a particular problem for post-surgical anatomy or unusual body habitus.
For clinical speed and simplicity, a single MRI sequence is generally sufficient for sCT prediction. For high-accuracy applications -- particularly in bone-rich regions or when precise dosimetric calculation is critical -- multiple complementary sequences (such as combining Dixon for tissue characterization with ZTE or UTE for bone detection) are preferred. Dixon MRI was specifically found to be adequate for quantitatively accurate sCTs, while ZTE improved air distribution accuracy in the abdomen.
Deep learning methods overall outperform traditional approaches in dosimetric accuracy, speed, and generalizability. However, they require large training datasets of paired MRI-CT scans, which are time-consuming to curate. GANs can also exhibit instability during training and may generate plausible-looking but physically incorrect HU values if not carefully validated. Current methods also do not fully replicate the radiomic feature distributions of real CT, suggesting they should not be used for CT-based radiomic analyses.
Standardization of validation procedures is identified as the most pressing need in the field. Studies use different evaluation metrics, anatomical sites, and patient cohorts, making direct comparison difficult. The field would benefit greatly from multi-center validation studies using standardized phantoms and clinical endpoints, moving beyond single-institution proof-of-concept work.
Remaining technical challenges include generating accurate sCT for patients with metal implants (which create severe artifacts in both real CT and MRI-derived estimates), handling respiratory and cardiac motion in the thorax and abdomen, and improving bone HU accuracy -- which is critical for proton therapy where small density errors cause large changes in the proton beam's stopping depth.
The integration of sCT with MRI-guided radiotherapy on hybrid machines like the MR-Linac represents the most ambitious future application: real-time imaging during radiation delivery, combined with AI-driven sCT generation, could enable daily adaptive treatment based on the tumor and anatomy seen moments before each treatment fraction. This trajectory points toward a future where MRI entirely replaces CT in the radiation oncology workflow.