KEY POINTS
- Investigators analyzed 100 transposed ovaries from 51 cervical cancer patients receiving postoperative VMAT to 45 Gy in 25 fractions and compared a log-linear distance model, random forest and 3D U-Net.
- Models were evaluated with patient-wise five-fold cross-validation. Random forest achieved the lowest overall error, with MAE 0.803 Gy and median absolute error 0.365 Gy, compared with MAE 0.886 Gy for log-linear modeling and 0.890 Gy for the 3D U-Net.
- At the clinically tight ±0.5-Gy tolerance, random forest correctly predicted 59% of ovarian doses compared with 41% for the log-linear model.
- Prediction performance depended strongly on geometry. Residual analysis identified a turning point at approximately 17.1 mm between the ovary and PTV.
- When ovary–PTV distance was ≥17 mm, random forest performed well, with MAE 0.390 Gy and Bland–Altman limits of agreement of approximately −0.97 to +0.98 Gy. Below 17 mm, MAE exceeded 1.8 Gy for all three models and limits of agreement widened markedly.
- For identifying ovaries receiving a mean dose above 6 Gy, random forest achieved an AUC of 0.956. However, the 17-mm threshold and all accuracy estimates came from a single-centre dataset without external validation.
CLINICAL TAKEAWAY
The most useful finding may be the failure boundary rather than the winning algorithm. When a transposed ovary lies within roughly 17 mm of the PTV, geometry becomes difficult enough that automated dose prediction should not substitute for manual planning review.