KEY POINTS
- The single-institution study used 220 segmented planning CT scans from 216 prostate cancer patients who had undergone elective pelvic nodal irradiation and multidisciplinary contour peer review between 2021 and 2024.
- The development dataset included 179 training and 41 internal-validation scans, with an additional 10 completely withheld prostate nodal cases used to test model performance independently of model selection.
- Most patients had pelvic nodal treatment to 45 Gy in 25 fractions or 50.4 Gy in 28 fractions, often with simultaneous prostate treatment to 70 Gy in 28 fractions. The model was based on a U-Net architecture with standardized image normalization and data augmentation.
- In internal validation, median Dice similarity coefficient was 0.79, median surface distance 1.68 mm, and Hausdorff distance 35.4 mm. In the completely withheld test cohort, Dice remained 0.79 (95% CI 0.71–0.86), median surface distance was 2.07 mm, and Hausdorff distance 27.1 mm.
- The apparently good global overlap did not eliminate clinically meaningful localized differences. Recurrent disagreements involved areas already recognized as difficult in consensus contouring, including the superior common iliac border, external iliac-to-inguinal transition, obturator region, and anterior nodal extent.
- These disagreement regions overlapped with areas of high interobserver variability reported in the NRG pelvic nodal consensus atlas, suggesting that the model could function as a standardized “second opinion” rather than simply an automatic contour generator.
- Generalizability is limited by the single-institution dataset, strong influence of a small number of physicians, and the fact that all reference contours had already undergone local peer review. Large target volumes may also make Dice scores look reassuring despite focal clinically important errors.
CLINICAL TAKEAWAY
The more interesting use of this model may be QA rather than automation: placing an AI-generated pelvic nodal contour beside the physician contour could immediately highlight areas worth discussing at peer review. A Dice of 0.79 is not sufficient to replace expert contouring, particularly at precisely the anatomical borders where consensus remains difficult.