MRI-sCT_converter-DoseRAD2026

MRI → Synthetic CT Converter — DoseRAD2026

A 3D regression model that generates synthetic CT (sCT) from MR images, trained for the DoseRAD2026 Grand Challenge proton dose on MRI task. It is the front end of team DoseHappens's MRI submission: the sCT it produces is what the protonand photon dose engine then computes on. The data was registered using the existing deformation of Boussot et al..

The model is trained with an nnUNet regression pipeline (3D full-resolution) and exported to TorchScript, so no nnUNet installation is required to run it. The training code is a modified fork of nnUNet.

The trainer is nnUNetTrainerRegression_impact: deep-supervision MAE with an added feature-space regularization term against the source MR (loss = MAE_deep + 0.1 · IMPACT(prediction, source)). It subclasses nnUNetTrainerRegression_mae_deep, so the IMPACT term is added to the MAE loss rather than replacing it. Both are in the fork linked above.

The dose engine and correction network that consume this sCT are in the challenge code repository: https://github.com/LFetty/doserad2026-maalgo

Model Files

model.pt is a TorchScript module exported with torch.jit.save() and must be loaded with torch.jit.load(). The folder is named mrtoct_bundle/ so it drops straight into the challenge container, which expects it at model_mri/mrtoct_bundle/.

Requirements

torch
scipy
numpy

Quick Start

import numpy as np
from standalone_regression_inference import StandaloneRegressionInference

model = StandaloneRegressionInference(
    model_path="mrtoct_bundle/",
    device="cuda",          # or "cpu"
    inference_amp=True,     # fp16 autocast for the network forwards
    patch_batch_size=2,     # patches per model call
)

# mr_volume: 3D numpy array, shape (D, H, W)
mr_volume = np.load("your_mr_volume.npy")

sct = np.asarray(model.predict(mr_volume)).squeeze()
sct = np.clip(sct, -1024, 3071)          # see "Important" below

predict() handles internally:

• Z-score normalization of the MR input
• Sliding-window tiled inference with Gaussian patch blending
• Denormalization of the sCT output back to HU with the precomputed statistics

Important: clip the output

predict() returns unclipped denormalized HU. The submitted pipeline clips to [-1024, 3071] before the sCT is used for density/material conversion. Without that clip, out-of-range predictions propagate into the dose calculation. The clip is applied outside this class, so you must apply it yourself.

Training Details

Limitations

On the DoseRAD2026 MRI task the synthetic CT is the accuracy bottleneck, not the dose engine. Improvements that gave ~12.5% on the CT task returned only ~2% through the MRI path, which places the remaining error in the MR→sCT step rather than in dose computation. Body-region CT MAE measured on one held-out case during precision verification was ≈24.75 HU; this is a single-case figure from a precision comparison, not a validation study.

Citation

If you use this model, please cite the DoseRAD2026 challenge and link this repository.