PG-Diff: a geometry-guided conditional diffusion model for dose prediction in cervical cancer brachytherapy

Dose distributions in cervical cancer brachytherapy are characterized by an extremely high dynamic range and steep dose gradients, making it challenging for conventional deep learning models to accurately reconstruct high-dose regions adjacent to applicators and dose boundaries at the interface between the target and organs at risk. In this study, we propose PG-Diff, a geometry-guided conditional diffusion model, for accurate three-dimensional dose distribution prediction in cervical cancer brachytherapy. This study included 188 patients with cervical cancer who underwent brachytherapy at three centers; the cohort was divided into training, validation, and test sets at a ratio of 7:1:2. The model used CT images, segmentation masks, and applicator prior information as conditional inputs, and formulated dose prediction as a conditional diffusion-based denoising generation process. To explicitly incorporate the physical principles of brachytherapy, we designed a physical feature fusion module, termed PhysicalFusion, which enables cross-modal interaction between anatomical semantic features and geometric distance features at the network bottleneck, thereby guiding the model to learn the nonlinear mapping between distance and dose intensity. Model performance was evaluated using D mean and D 90% for the high-risk clinical target volume (HR-CTV), D 1cc and D 2cc for the bladder, colon, rectum, and small bowel, and spatial metrics (MAE, E95, and G-RMSE). PG-Diff achieved the best predictive performance across most dosimetric metrics. For HR-CTV, the prediction errors of D mean and D 90% achieved by PG-Diff were 0.73 ± 0.41 Gy and 0.29 ± 0.22 Gy, respectively, both lower than those of the comparison models. For organs at risk, PG-Diff yielded lower errors in predicting D 1cc and D 2cc for the colon, rectum, and small bowel; notably, the D 1cc and D 2cc errors for the small bowel were 0.14 ± 0.17 Gy and 0.11 ± 0.13 Gy, respectively. The MAE, E95, and G-RMSE were 2.06 ± 0.69 Gy, 7.89 ± 2.80 Gy, and 2.50 ± 0.61 Gy/mm, respectively. Visual assessment and dose–volume histogram analysis showed that the dose distributions generated by PG-Diff were closer to the ground-truth planned doses, particularly demonstrating improved spatial consistency in high-dose regions adjacent to applicators and in HR-CTV dose coverage regions. The proposed PG-Diff effectively improves the accuracy and physical consistency of three-dimensional dose prediction for cervical cancer brachytherapy, enabling more precise characterization of high-gradient dose regions and low-dose distributions in organs at risk. This method provides a clinically translatable approach for automated and rapid generation of brachytherapy treatment plans.