KEY POINTS
- This French multi-institutional study evaluated four nnU-Net training strategies for automated cervical cancer segmentation on cone-beam CT (CBCT). The full dataset comprised 73 patients from seven institutions, with 23 patients used for initial single-center model training and 50 patients available for model adaptation and evaluation. After excluding seven patients used for population-level fine-tuning, the final evaluation included 430 CBCT scans from 43 patients.
- Four strategies were compared: a baseline nnU-Net trained on single-institution CT and CBCT data; joint training with CBCT and synthetic CT (sCT); population-level fine-tuning using data from seven patients; and patient-specific fine-tuning using one manually contoured planning CT per patient. The evaluated structures were the tumoral clinical target volume (CTVt), bladder, and rectum, with segmentation quality classified by an expert radiation oncologist as good, acceptable, or unacceptable.
- Patient-specific fine-tuning substantially improved target segmentation quality. The proportion of CTVt segmentations rated good increased from 23.7% (102/430) with the baseline model to 55.8% (240/430) after patient-specific fine-tuning, an absolute improvement of 32.1 percentage points. Corresponding results were 24.4% with CBCT plus sCT training and 24.0% with population-level fine-tuning, indicating little benefit from these alternative strategies.
- Bladder and rectum segmentation improved substantially with individualized adaptation. Good bladder segmentations increased from 29.1% to 49.1%, while good rectum segmentations increased from 33.0% to 64.4%. The proportion of unacceptable CTVt segmentations decreased from 21.4% to 10.5%. Most improvements represented transitions between adjacent quality categories rather than changes from completely unacceptable to immediately usable contours.
- Patient-level statistical analysis confirmed the improvement. After aggregating repeated CBCT assessments within each patient, fine-tuning significantly improved expert ratings for CTVt (p<0.001), bladder (p=0.002), and rectum (p=0.001). Neither CBCT plus sCT training nor population-level fine-tuning demonstrated a significant improvement in this analysis. Performance varied between institutions, particularly those with smaller patient numbers or different imaging characteristics.
- Individualized adaptation required approximately one additional hour per patient, using an already available manually contoured planning CT and 10 training epochs. Baseline model training required approximately two days, while population-level fine-tuning required an additional 12 hours. Inference took approximately two minutes per CBCT for all strategies. The study therefore demonstrated technical feasibility for pretreatment model adaptation, although complete clinical workflow efficiency was not evaluated.
- The principal limitation was the absence of independent manual CBCT reference segmentations. Model quality was assessed by one expert, and predictions with Dice similarity >0.90 to baseline for all three structures generally retained their original ratings without complete independent reinspection. Consequently, the results demonstrate improved expert-rated usability rather than objectively validated contouring accuracy. No dosimetric comparison, treatment adaptation outcome, or reduction in clinical workload was demonstrated.
CLINICAL TAKEAWAY
Patient-specific fine-tuning substantially improved expert-rated cervical cancer CBCT segmentation across institutions using only a previously contoured planning CT. This approach may be more effective than population-level adaptation or synthetic CT augmentation when implementing automated segmentation across different imaging systems. However, independent manual-contour validation and workflow studies are required before its accuracy and clinical utility can be established.