AI-generated synthetic CT shows promising dose accuracy in prostate patients with hip implants

In 20 prostate patients with hip replacements, implant-trained synthetic CT models closely reproduced CT dose, but two cases with rectal air had outliers.

KEY POINTS

  • This retrospective study from University Hospital Zurich, Switzerland, investigated deep learning-based synthetic CT (sCT) generation for MRI-only radiotherapy planning in 20 patients with prostate cancer and total hip replacements. Patient-level five-fold cross-validation was performed, with 16 training and four test patients per fold, preserving unilateral and bilateral implant distributions. An additional 16 implant-free patients were included in a separate mixed-training experiment to determine whether broader training data improved sCT quality in patients with metal implants.
  • The approach used a Pix2Pix conditional generative adversarial network trained on paired MRI and planning CT images. Input data consisted of 1.5-T T1-weighted Dixon MRI, using in-phase and opposed-phase reconstructions, while registered reference CT images underwent metal artifact reduction. The model was trained on two-dimensional axial slices to generate tissue-density information, including high-density hip implants despite signal loss and distortion on MRI. Metal-containing regions were assessed separately using CT-derived implant masks and image-similarity metrics.
  • Dosimetric validation was performed using simulated prostate IMRT plans prescribing 40 Gy in five fractions. Plans were optimized on sCT and recalculated on reference CT using the matRad pencil-beam algorithm with a 2-mm dose calculation grid. Both implant-avoiding and more challenging beam geometries traversing the prostheses were evaluated. Target metrics included PTV D95%, D2%, and mean dose, while bladder and rectal dose-volume parameters were compared. No density overrides were applied to sCT, although reference CT calculations included corrections for relevant rectal air differences.
  • The best-performing investigated configuration achieved close image and dose agreement. Mean absolute error was 71.62 HU across the body contour, 280.20 HU in the implant region, and 49.30 HU outside the implant region. Representative mean DVH differences were below 0.5%, including approximately 0.02% for PTV D95% in the partial-arc scenario. Mean 1%/1 mm gamma passing rate was 98.37% ± 2.95% using a 90% dose threshold, while mean 2%/2 mm gamma agreement reached 100% at both evaluated dose thresholds.
  • Most cross-validation folds demonstrated stable dosimetric agreement, but clinically relevant outliers occurred. Four of five folds had no DVH deviations exceeding the predefined ±2% criterion, whereas two patients in the fourth fold accounted for the observed outliers. Both had substantial rectal air inclusions not represented in the training cases, producing errors in synthetic CT generation and subsequent dose calculation. These findings demonstrate that good average image similarity and gamma agreement do not exclude clinically important patient-specific failures.
  • Training on implant-containing images was essential for reliable metal reconstruction. A model trained exclusively on 16 patients without implants failed to reconstruct the hip prostheses correctly in an exploratory bilateral-implant case, generating false air regions and dosimetric outliers when beams traversed the implants. Adding implant-free images to the implant-containing training set reduced variability across cross-validation folds, but no statistically significant improvement in DVH accuracy was demonstrated compared with implant-only training. Thus, additional training data were not uniformly beneficial without adequate representation of the clinically challenging anatomy.
  • Several safety-critical elements of an MRI-only clinical workflow were not validated. The study included only 20 implant patients from one institution and MRI scanner, with no independent external test cohort. Implant materials and manufacturers were not evaluated separately, and the dose calculations used a pencil-beam algorithm rather than independent high-accuracy Monte Carlo verification. The investigation did not assess geometric MRI distortion around metal, synthetic CT-to-CBCT registration accuracy, or actual treatment delivery. Consequently, the results establish technical feasibility rather than readiness for routine CT-free treatment.

CLINICAL TAKEAWAY

Deep learning-based synthetic CT generation may extend MRI-only prostate radiotherapy planning to selected patients with hip replacements, a population often excluded from existing workflows. However, failures associated with unexpected rectal air and models trained without metal-containing anatomy demonstrate the need for representative training data and patient-specific verification. Independent dosimetric and geometric validation remains necessary before safely replacing planning CT in these patients.

SOURCE

Physics and Imaging in Radiation Oncology

Browse more research Suggest a correction