ESTRO 2024 - Abstract Book

S160

Brachytherapy - Breast

ESTRO 2024

the fixation buttons of the catheters as landmarks. To enhance the reconstruction accuracy beyond the intrinsic accuracy of the EMT system, the EMT implant data were refined applying methods of image segmentation. For this, the EMT reconstruction points were used as a prior to define small regions of interest (ROI) around them within the CT image. The catheter track within these ROIs was segmented using thresholding and blob detection. The quality of the refined implant reconstruction was evaluated based on the geometric agreement with the clinically used, manual reconstruction performed by an experienced treatment planner according to GEC-ESTRO guidelines [4], which served as a ground truth. In addition, the time required for EMT data acquisition, processing and refinement was examined. A graphical user interface was designed to enable in-room visualisation and patient-specific tuning of the refinement parameters.

Results:

The refined EMT-based reconstruction of the catheter implant showed very high agreement with the clinical, manual reconstruction. Although the patient cohort included a broad spectrum of patients regarding age (37-83 years), implant size (11-27 catheters, PTV 18-215 cm 3 ), and breast size (130-1370 cm 3 ), the median perpendicular distance between the reconstructions across all patients was 0.24 mm, with a 95 th percentile of 1.1 mm. In 71 out of the 72 investigated patients, the median accuracy of the automated reconstruction was lower than the human inter-observer variability of 0.60±0.35 mm [5]. Only in one exceptional case, the approach was not successful because the EMT implant reconstruction could not be mapped to the CT image space due to unusually large deformations. Figure 1 shows an exemplary reconstruction before and after refinement for visual comparison. EMT data acquisition took between 5.5 min and 13 min, depending on the number of catheters, with an average of 8.6 min. Refinement took around 0.5 s per reconstruction point and usually required no more than 3 min per implant (range 0.5-5.5 min) without user intervention.

Conclusion:

The presented hybrid approach for automated implant reconstruction using EMT and complemented by image segmentation techniques demonstrated excellent accuracy and speed compared to manual reconstruction. In