S144
Brachytherapy - Physics
ESTRO 2026
10.1016/j.ctro.2018.01.001. PMID: 29594251; PMCID: PMC5862686.
Digital Poster 4505 Automated Needle Reconstruction in HDR Prostate Brachytherapy Ryan T Duffy Radiotherapy Physics, University Hospitals Sussex, Brighton, United Kingdom Purpose/Objective: HDR prostate brachytherapy delivers a radioactive source through transperineally-inserted needles, allowing dose escalation while limiting toxicity to surrounding tissue. Patient movement between implantation and treatment (typically 4-5 hours) can lead to needle displacement exceeding 5 mm in 67% of cases[1], compromising dosimetric accuracy. Currently, displacement assessment is performed manually by clinical scientists, which is time- consuming and subject to interobserver variation. Automatic segmentation of the needles is challenging due to the presence of confounding structure, close needle proximity and crossovers. Previous approaches have focussed on deep learning applications[2-4], however these use small samples from a single centre, limiting their robustness. This study aims to develop and validate an automatic method for detecting and reconstructing needle positions on paired CT datasets, providing a foundation for the reliable assessment of needle displacement between planning and treatment images. Material/Methods: CT images were cropped to retain the central 50% of pixels and thresholded to retain those voxels ≥ 800 HU. The negative distance transform of this volume was processed using watershed segmentation, with the coordinates of each segment’s maximum intensity pixel used to represent needle locations. Initial needle positions were labelled using the clinical insertion template and matched to segmented structures in the following slice with a Euclidean cost function. Needle trajectories were then reconstructed using a linear Kalman filter, which propagated position estimates slice-by-slice while modelling uncertainty. Segment associations were determined using the Mahalanobis distance between the predicted needle position and observed structure, incorporating a combination of greedy matching for strong associations and linear sum assignment for conflict resolution. Results: The method was evaluated on 11 anonymised patient datasets, totalling 22 CT scans. Across all cases, 99.5% of needles were correctly digitised, with accurate reconstruction achieved for both simple and more
Creation and export of the AI contours can be completed in ~3 minutes, whereas the manual contouring workflow takes on average 40±13 minutes per patient. Introducing the AI contours as a starting point is expected to significantly reduce this, and a study is ongoing to validate this. This time saving will directly impact the rate at which patients can start their treatment. Conclusion: This work demonstrates that an in-house solution for autocontouring can reliably produce contours on MR scans with variable tilt angles.The contours produced are of a sufficient quality to provide efficiency improvements in a clinical workflow alongside additional manual adjustments and is a significant improvement over reviewed commercial solutions. These results are a vital step towards ensuring that all patient groups can benefit from the introduction of AI automation within radiotherapy. Keywords: AI, Autocontouring, Automation References: [1] Dimopoulos, J et al. (2012). Recommendations from Gynaecological (GYN) GEC-ESTRO Working Group (IV). Radiotherapy and oncology : journal of the European Society for Therapeutic Radiology and Oncology, 103(1), 113–122. https://doi.org/10.1016/j.radonc.2011.12.024[2] Isensee, F. et al (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2), 203-211.[3] Pötter R et al. EMBRACE Collaborative Group. The EMBRACE II study: The outcome and prospect of two decades of evolution within the GEC-ESTRO GYN working group and the EMBRACE studies. Clin Transl Radiat Oncol. 2018 Jan 11;9:48-60. doi:
Made with FlippingBook - Share PDF online