Development of an Artificial Intelligence-Driven Three-Dimensional Reconstruction Model of Renal Vasculature for Surgical Planning in Robot-Assisted Partial Nephrectomy.

BACKGROUND

This single-centre study aimed to implement and internally validate a deep learning pipeline based on a 3D SegResNet architecture for automatic segmentation of renal vascular structures from contrast-enhanced CT images, enabling accurate three-dimensional reconstructions to support surgical planning in robot-assisted partial nephrectomy.

METHODS

Ninety-seven CT scans from patients with renal masses treated between January 2019 and September 2025 were included, with 68 used for training and 29 held out as an independent test set, never seen by the model during training. Imaging data included arterial and portal phases, and manual segmentations served as the ground truth. A 3D SegResNet architecture was implemented with spatial normalisation, isotropic resampling, data augmentation and tailored pre- and post-processing.

RESULTS

The model achieved mean DICE scores of 0.86 for renal arteries and 0.81 for renal veins, with HD95 of 6.4 ± 2.8 mm and ASSD of 0.58 ± 0.22 mm for arteries, and HD95 of 8.7 ± 3.5 mm and ASSD of 0.94 ± 0.41 mm for veins, demonstrating high concordance with manual segmentation and correctly identifying most vascular structures. Performance was lower for distal intraparenchymal branches and in cases with multiple renal arteries or veins.

CONCLUSIONS

Overall, this feasibility study supports the technical viability of automated, standardized and cost-effective tools for surgical planning.