Radiation therapy (RT) for prostate cancer requires precise calculation of how the radiation dose will distribute through the body. This calculation depends on knowing the electron density of different tissues -- essentially how much each tissue will absorb or allow radiation to pass through. CT scans provide this information through a measurement called Hounsfield units (HU).
However, CT images have poor soft tissue contrast, making it difficult to precisely outline the prostate and nearby structures. MRI provides far superior soft tissue visualization, which is why both scans are typically performed: CT for dose calculation and MRI for target delineation.
Using two separate imaging sessions creates real problems. The patient must maintain the same position across both scans, but small positional differences between CT and MRI sessions introduce registration errors that can shift tumor outlines by millimeters. This translates directly into uncertainty in exactly where the radiation dose is being aimed.
An MRI-only simulation workflow -- eliminating the CT scan entirely -- would simplify the process, reduce patient burden, and remove registration errors. The critical barrier is that MRI alone cannot directly provide the HU values needed for dose calculation.
Synthetic CT (sCT) refers to a computationally generated CT-like image derived from an MRI scan. If an AI model can learn the relationship between MRI signal patterns and CT Hounsfield values from training data, it can produce a virtual CT image without requiring the patient to undergo an actual CT scan.
Early deep learning approaches used Generative Adversarial Networks (GANs) -- a framework where one neural network (the generator) creates synthetic images and another (the discriminator) evaluates their realism. This back-and-forth training competition drives increasingly accurate image synthesis.
Standard GAN training requires paired images -- matched CT and MRI scans from the same patient in the same position. In practice, even same-day scans are imperfectly registered due to patient movement and organ shifting, meaning paired training data is never perfectly aligned.
A more flexible approach, Cycle-Consistent GAN (CycGAN), was developed to handle unpaired data. It trains two generators simultaneously: one converting MRI to CT, and one converting CT back to MRI. The requirement that these round-trip conversions are self-consistent (cycle-consistent) enables training even without perfectly matched image pairs.
The study used paired CT and MRI data from 113 prostate cancer patients who underwent radiation therapy at Samsung Medical Center. Scans were performed on the same day with patients in the same position, and the dataset was split into 93 training/validation cases and 20 test cases.
Three deep learning models were compared. The first was a standard GAN trained with paired images using mean absolute error (MAE) to minimize differences between the generated and actual CT. The second was CycGAN, trained without requiring perfectly paired images using cycle-consistency loss. The third, RgGAN (Reference-guided CycGAN), combined CycGAN's unpaired flexibility with additional MAE supervision from the available paired reference images.
To evaluate image quality, Hounsfield unit accuracy was measured for different tissue types (body, bone, soft tissue). Dice similarity coefficients compared the structural overlap between synthetic and real CT images, and Hausdorff distances measured the maximum boundary displacement -- both metrics reflecting how closely the synthetic image matches the real one anatomically.
The critical clinical test was dosimetric evaluation. Radiation treatment plans originally calculated on real CT scans were recalculated on the synthetic CTs, and the resulting dose distributions were compared. If an sCT can reproduce the planned dose within 2% for the tumor and surrounding organs, it is considered clinically acceptable for treatment planning.
GAN produced sCTs that most closely matched the geometric structure of the real CT images -- it had lower mean absolute error and higher Dice similarity coefficients for bone and soft tissue than CycGAN. This is expected since GAN trains directly on paired images and explicitly minimizes pixel-by-pixel differences.
However, GAN showed a significant weakness in Hounsfield unit accuracy for soft tissue. Only the GAN-generated images had a statistically significant difference in mean HU from the real CT for soft tissue (p = 0.03). CycGAN and RgGAN preserved soft tissue HU values more faithfully -- a critical property for accurate dose calculation since dose depends on tissue density.
CycGAN had the worst structural similarity in terms of Dice and Hausdorff metrics for both bone and soft tissue. This likely reflects the fact that CycGAN does not directly compare each generated pixel against the paired real CT -- it only enforces round-trip self-consistency, which does not guarantee geometric fidelity.
RgGAN bridged the gap between the approaches. By adding paired reference supervision to the CycGAN framework, it preserved better HU accuracy for soft tissue than GAN, while improving geometric similarity over pure CycGAN. This suggests the hybrid approach represents a meaningful advance over either method alone.
All three deep learning methods produced sCTs suitable for clinical radiation treatment planning. Mean relative dose differences for all planning target volumes and organs at risk were within 2% compared to the original CT plans -- the standard clinical threshold for acceptability.
However, there were meaningful differences in specific dose metrics. The D98% and D95% values -- representing the dose received by 98% and 95% of the tumor volume, the most clinically critical coverage metrics -- showed significantly larger deviations from the real CT when using GAN-based sCTs compared to CycGAN or RgGAN.
Between CycGAN and RgGAN, RgGAN showed superior performance for D98% and D95%, although the difference did not reach statistical significance due to the small test set size. This trend supports the hypothesis that better soft tissue HU accuracy translates into more accurate dose coverage of the tumor.
Gamma pass rates -- a composite measure of dose spatial accuracy -- were consistently lowest for CycGAN and significantly worse than RgGAN for the stricter tolerance criteria. For organs at risk including the bladder, rectum, and femoral heads, no significant dosimetric differences were found between any of the three methods.
A key insight from this study is that geometric similarity to the original CT does not directly predict dosimetric accuracy. GAN had the best geometric metrics but worse dosimetric performance for high-dose tumor coverage. Soft tissue HU accuracy proved to be the more relevant predictor of radiation dose accuracy -- a finding with important implications for how sCT models should be designed and evaluated.
The use of T2-weighted spin echo MRI in this study is clinically practical but carries image blurring due to signal decay effects. Higher quality or additional MRI sequences (such as Dixon water-fat imaging or zero echo time sequences for bone) could improve sCT generation, particularly for accurately capturing bone density.
Important limitations include the small 20-patient test set, which limits statistical power, and the single-institution design. The models were trained at one hospital with specific imaging protocols, and performance may differ when applied to scans acquired with different MRI sequences, field strengths, or patient preparation protocols at other centers.
The RgGAN model's approach -- combining the flexibility of unpaired training with the precision of paired reference supervision -- represents a promising design philosophy that could be further refined. Future work using larger datasets, multiple MRI sequences, and prospective clinical validation would be needed before routine clinical deployment.