KEY POINTS
- The study evaluated a commercial MRI-only workflow using MVision Contour+ and Image+ in 19 male pelvis datasets from the multi-institutional Gold Atlas. Five experts had independently delineated the bladder, rectum, prostate, seminal vesicles and penile bulb, allowing AI contours to be compared with both individual observers and a STAPLE consensus reference.
- AI segmentation generally remained within manual interobserver variability. Mean Dice coefficients were 0.92 for bladder, 0.90 for prostate, 0.86 for rectum, 0.76 for seminal vesicles and 0.72 for penile bulb; the leave-one-out kappa analysis did not identify the AI segmentation as systematically distinguishable from the five human observers.
- Several statistically significant systematic contour differences were nevertheless present. AI prostate volume was a median 5.2% smaller, seminal vesicle volume 13.8% smaller, and penile-bulb volume 15.3% larger than the mean manual result; bladder Dice was lower by 0.029. These shifts were generally small but show that automated contours still require clinical review.
- Synthetic CT image similarity was high: median mean absolute error was 34.0 HU, structural similarity index 0.93, peak signal-to-noise ratio 29.8 dB, and normalized cross-correlation approximately 0.89. The largest differences occurred around the body surface and high-density bone rather than pelvic soft tissue.
- For dosimetric testing, each patient received a dual-arc VMAT plan prescribing 42.7 Gy in seven fractions to a prostate PTV generated with a 5-mm margin. The exact beam parameters were transferred from conventional CT to synthetic CT so that the comparison isolated differences in electron-density representation.
- Dose agreement was extremely close: median three-dimensional gamma pass rate was 99.8% at 2%/2 mm and 99.5% at 1%/1 mm. Median mean absolute dose difference was only 0.036 Gy, 95.9% of voxels were within ±0.1 Gy, and the Dice coefficients for the 95% and 50% isodose volumes were 0.996 and 0.997.
- All evaluated clinical target and OAR constraints remained satisfied after synthetic-CT recalculation. Generalizability remains limited by the 19-patient sample, use of an existing atlas rather than prospective clinical workflow, evolving contouring conventions and the fact that actual time savings or error reduction from eliminating CT-MRI registration were not measured.
CLINICAL TAKEAWAY
A commercial AI pipeline combining auto-segmentation and synthetic CT produced prostate-planning geometry and dose calculation very close to expert and CT-based references. The data support clinical implementation studies of MRI-only planning, but do not support unsupervised contour acceptance, particularly for smaller or less consistently defined structures.