KEY POINTS
- This retrospective study included 2,394 single-fraction prostate HDR brachytherapy plans, all prescribing 15 Gy, with 2,155 plans used for five-fold model development and 239 completely held out for testing.
- A 3D U-Net received only the contoured prostate, urethra, rectum and possible dwell locations; actual dwell times were deliberately excluded. The model predicted the expected three-dimensional 100% prescription isodose volume for comparison with the approved clinical plan.
- Prostate coverage prediction was highly accurate. On the held-out test set, the mean absolute V100 error was 0.94 percentage points, with 88.3% of cases within ±2 percentage points and 99.6% within ±5.
- Predicted and clinical 100% isodose volumes showed almost one-to-one volumetric agreement: R² 0.990, regression slope 0.99 and median absolute volume difference 1.04 cm³, corresponding to a 2.5% median relative difference.
- Spatial agreement was also high, with a median Dice coefficient of 0.960, median average symmetric surface distance of 0.48 mm and median HD99 of 1.0 mm. Every test case achieved Dice ≥0.90.
- Validation-derived thresholds were designed to identify unusual plans rather than declare errors. At the 95% coverage-error threshold, 6.3% of test plans were flagged for targeted review; spatial thresholds identified similarly small subsets with atypical Dice or surface-distance values.
- The model learns institutional planning patterns rather than independently recalculating TG-43 dose. It was trained exclusively on 15 Gy whole-gland plans, so focal boosts, salvage brachytherapy, other fractionations and other institutions require dedicated validation or retraining.
CLINICAL TAKEAWAY
Traditional brachytherapy nomograms can tell a physicist that the total dwell time looks unusual, but they cannot show where the plan is unusual. This 3D approach adds a potentially useful spatial consistency check before treatment, but it should complement established HDR QA rather than replace independent dose verification.