S84
Brachytherapy - Gynaecology
ESTRO 2026
brachytherapy with Venezia or Geneva applicators showed a median dwell position difference with clinical manual reconstruction of only 0.66 mm. The total time needed for AI-based applicator reconstruction was less than 2 minutes, including image processing. This is a significant potential reduction in cervical cancer brachytherapy. Keywords: applicator reconstruction, AI, cervical cancer References: [1] van Vliet-Pérez SM, van Paassen R, Wauben LSGL, et al. Time-action and patient experience analyses of locally advanced cervical cancer brachytherapy. Brachytherapy. 2024;23(3):274-281. [2] Isensee F, Jaeger PF, Kohl SAA, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18:203-211.[3] Klein S, Staring M, Murphy K, Viergever MA, Pluim JPW. elastix: a toolbox for intensity based medical image registration. IEEE Trans Med Imaging. 2010;29(1):196- 205. Mini-Oral 4804 ctDNA-Guided Personalization of 3D-Printed Applicators for Enhanced Brachytherapy Planning in Locally Advanced Cervical Cancer Wiwatchai Sittiwong 1,2 , Yusung Kim 2 , Lauren Colbert 2 , Shiqin Su 2 , Tatiana Cisneros Napravnik 2 , Kyoko Court 2 , Pittaya Dankulchai 1 , Tissana Prasartseree 1,2 , Ann Klopp 2 1 Division of Radiation Oncology, Department of Radiology, Faculty of Medicine Siriraj hospital Mahidol University, Bangkok, Thailand. 2 Radiation Oncology, MD Anderson Cancer Center, Houston, Texas, USA Purpose/Objective: To assess whether early changes in circulating tumor DNA (ctDNA) dynamics within the first 2 weeks of concurrent chemoradiotherapy (CCRT) can predict the potential benefit of patient-specific 3D-printed applicators in 3D image-guided adaptive brachytherapy (3D-IGABT) for locally advanced cervical cancer (LACC). Material/Methods: Twenty LACC patients from a prospective ctDNA monitoring cohort who underwent CCRT and 3D- IGABT were included. For each patient, an additional brachytherapy plan was retrospectively generated using a patient-specific 3D-printed applicator template to enhance HR-CTV coverage while maintaining or improving organ of interest doses, and dose–volume histogram parameters were compared with the original clinical plan. A composite dosimetric benefit score was calculated based on improvements in target coverage and organ of interest sparing. For HR-CTV
automated applicator reconstruction for CT and MR and applicator types used in our clinic. Material/Methods: Clinical images with applicator reconstructions from 354 treatment fractions of 173 patients were included, including MRI and CT images and all applicator types (Venezia, Geneva) and sizes used in the clinic. Data were stratified for training/testing: 105/27 with MR+Venezia, 40/9 with MR+Geneva, 101/25 with CT+Venezia and 39/8 with CT+Geneva. AI-based automated reconstruction consisted of two steps i) nnU-Net [2] based automated segmentation, followed by ii) automated registration of segmented and library applicator models with the ITK-Elastix toolbox [3]. Three applicator reconstruction workflows were developed: separate training with MR or CT (MRmodel and CTmodel, respectively), and using both MR and CT for multi-modality training (MMmodel). Training always included images with Venezia and with Geneva. Dwell Position Differences (DPD) between clinical manual reconstructions and automated reconstructions were primarily used to evaluate reconstruction accuracy. Automated reconstruction times were evaluated. Results: Figure 1 shows very small DPDs for automated reconstruction with MMmodel, when tested on both MR and CT: median (IQR) DPD of 0.66 (0.50-0.81) mm. For MR test fractions only/CT test fractions only, DPD differences between MMmodel and MRmodel/CTmodel were clinically irrelevant (Figure 1). Differences in DPD between Venezia and Geneva were non-significant (p>0.05) and clinically negligible in all workflows. Three outlier fractions were found out of 69 test fractions (Figure 1), all corresponding to patients with disconnected ovoids. For the MMmodel workflow, automated reconstruction times were 110±53 s (including image loading and processing), which was slightly longer than for MRmodel (81±12 s) and CTmodel (76±13 s).
Conclusion: A single Deep Learning-based automated applicator reconstruction for MR- and CT-based cervical cancer
Made with FlippingBook - Share PDF online