SynthRAD2023

The primary task focuses on generating synthetic CT (sCT) scans from multi-contrast MRI inputs of the brain. Participants must accurately predict Hounsfield Units (HU) to enable precise radiotherapy dose calculations.

• Input: T1-weighted and T2-weighted MRI sequences
• Output: A 3D synthetic CT volume with accurate bone/soft-tissue contrast
• Clinical Goal: Eliminate the need for a separate planning CT scan, reducing patient radiation exposure and registration errors between MRI and CT.

This task extends the challenge to the pelvic region, where anatomical variability and organ motion present significant difficulties. Models must handle complex bone structures and soft tissue boundaries.

• Challenge: Higher anatomical variation and presence of gas pockets causing signal voids
• Metric Focus: Geometric accuracy of femoral heads and sacrum for robust patient alignment
• Impact: Validates the generalizability of synthesis algorithms across different anatomical sites.

Moving beyond MRI, this task evaluates the correction of Cone-Beam CT (CBCT) artifacts to generate high-quality planning CT scans. CBCT is widely used for patient positioning but suffers from severe scatter and noise.

• Input: Low-quality, noisy CBCT with limited field-of-view
• Output: Corrected, high-quality synthetic CT suitable for adaptive radiotherapy
• Key Metric: Preservation of electron density information for dose re-calculation during treatment.

SynthRAD2023 mandates a multi-faceted evaluation protocol to prevent over-optimization on a single metric. Rankings are determined by a composite score.

• MAE (Mean Absolute Error): Quantifies voxel-wise HU differences.
• PSNR (Peak Signal-to-Noise Ratio): Measures reconstruction quality.
• SSIM (Structural Similarity Index): Evaluates perceived structural degradation.
• Dose Accuracy: Gamma analysis comparing dose distributions calculated on sCT vs. real CT.

The challenge provides a large, multi-center dataset of paired MRI-CT and CBCT-CT scans to ensure statistically significant benchmarking. All data is de-identified and shared under a research license.

• Training Data: Includes ground-truth paired scans for supervised learning.
• Hidden Test Set: Final rankings are determined on a sequestered test set to prevent leaderboard hacking.
• Code Submission: Participants often submit containerized algorithms to ensure reproducibility and fair compute comparison.

The ultimate benchmark is clinical utility. SynthRAD2023 emphasizes that geometric similarity is insufficient; the synthetic output must be dosimetrically equivalent to a real CT.

• Electron Density Mapping: sCT must accurately map HU values to relative electron density for the treatment planning system.
• Gamma Pass Rate: The percentage of voxels where the dose difference and distance-to-agreement meet clinical tolerance (e.g., 2%/2mm).
• Workflow Integration: The winning solutions demonstrate seamless integration into existing radiotherapy pipelines.

The foundational task of mapping an input image from a source domain (MRI) to a corresponding output image in a target domain (CT). Unlike simple style transfer, this requires learning a bijective mapping that preserves anatomical geometry while accurately synthesizing tissue-specific Hounsfield Unit (HU) values. Architectures often employ encoder-decoder structures with skip connections to retain high-resolution structural details during the domain shift.

The quantitative standard for radiodensity in CT imaging, where water is defined as 0 HU and air as -1000 HU. A synthetic CT generator must not only produce visually convincing images but also assign precise HU values to bone, soft tissue, and air cavities. Errors in HU estimation directly corrupt dose calculation algorithms in radiotherapy planning, making this the most clinically critical evaluation metric in SynthRAD2023.

A perceptual metric that quantifies the degradation of structural information between a synthetic CT and the ground truth. Unlike simple pixel-wise error, SSIM evaluates luminance, contrast, and structure independently, mimicking human visual perception. It is a core evaluation metric in SynthRAD2023, ensuring that generated images preserve fine anatomical boundaries and do not introduce geometric distortions that could compromise treatment margins.

A distribution-level metric that measures the similarity between the feature statistics of generated synthetic CTs and real CTs. A lower FID score indicates that the synthetic images capture the textural diversity and realism of true CT scans. SynthRAD2023 uses FID to penalize mode collapse, where a generator produces only a limited variety of anatomically plausible outputs, failing to represent the full patient population.

The ultimate clinical endpoint for SynthRAD2023. Synthetic CTs are fed into Monte Carlo simulation or clinical treatment planning systems to compute radiation dose distributions. The challenge evaluates the gamma pass rate between dose maps calculated on synthetic CTs versus real CTs. A passing rate above 95% under strict criteria (e.g., 1%/1mm) is the gold standard for clinical acceptability in MR-only radiotherapy workflows.