Geometry-Aware Automated Correction of Brachytherapy Needle Trajectories In 3D Transrectal Ultrasound
Abstract
Purpose
Accurate localization of brachytherapy needles in transrectal ultrasound (TRUS) is essential for high-quality treatment planning, yet remains challenging due to speckle noise, needle shadowing, and variable image quality. This work presents a robust, geometry-aware method for automated correction of needle trajectories in 3D TRUS volumes, emphasizing accurate in-plane (XY) alignment across slices and reliable needle identification.
Methods
A deep learning framework was developed to refine initial needle trajectories using 3D TRUS volumes and per-needle geometric inputs. The model jointly regresses needle entry and tip points and incorporates an auxiliary needle mask head to provide voxel-level guidance. Five-fold cross-validation was performed on 1,035 TRUS volumes (932 training, 103 testing), comprising 12,818 training and 1,413 test needles. Because clinical initializations were unavailable retrospectively, initial needles were simulated from ground truth using random perturbations in position and orientation. Final inference used an ensemble of cross-validation models with test-time augmentation, followed by geometry-based post-processing guided by predicted needle voxels. Performance was evaluated using slice-wise in-plane distance metrics, three-dimensional entry and tip errors, and percentile-based statistics. Robustness to initialization quality was assessed via a sensitivity analysis with increasing in-plane perturbations.
Results
On the held-out test set, the proposed pipeline achieved 100% needle recall, successfully matching and correcting all needles. Median per-needle in-plane slice error after post-processing was 0.43 mm (IQR: 0.28-0.63 mm), with 95% of needle points within 1.1 mm. Per-case averaged median 3D tip and entry errors were 0.59 mm (IQR: 0.47-0.69 mm) and 0.50 mm (IQR: 0.44-0.63 mm), respectively. Per-case mean in-plane slice error was reduced from 2.64 mm to 0.47 mm. Performance remained stable for initialization offsets up to 5.0 mm.
Conclusion
This work demonstrates a robust, geometry-focused approach for automated needle correction in 3D TRUS, achieving complete needle recovery with clinically relevant geometric accuracy and supporting automation for brachytherapy workflows.