Click here to

Session: Therapy General ePoster Viewing [Return to Session]

Deep-Learning-Based Automated HDR Applicator Digitization for GYN Brachytherapy

S Momin1*, Y Lei2, T Wang3, M Axente4, J Roper5, J Shelton6, J Bradley7, T Liu8, X Yang9, (1) Emory Univ, Decatur, GA, (2) Emory University, Decatur, GA, (3) Emory University, Atlanta, GA, (4) Emory University, Atlanta, GA, (5) Winship Cancer Institute of Emory University, Atlanta, GA, (6) Emory University, ,,(7) Emory University School of Medicine, ,,(8) Emory University School of Medicine, Atlanta, GA, (9) Emory University School of Medicine, Atlanta, GA

Presentations

PO-GePV-T-36 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Manual digitization of the HDR GYN applicator can be time consuming when the needles are within close proximity of each other. The purpose of this study was to develop a deep learning-based method for automatic digitization of applicator and needles from computed tomography (CT) scans of female patients.

Methods: Our proposed method, named objective activation network, consists of three subnetworks, i.e., feature extractor, a one stage object detector, and a mask module. Feature extractor is used to extract informative features from CT. A one stage object detector is utilized to locate the volume-of-interest (VOIs) of applicators. Finally, mask module is then used to digitize applicators within the VOIs. In order to improve spatial consistency of applicators’ digitization, a total variation regularization term is used. Three-fold cross-validation is used for evaluating a cohort of 21 patients. The 21 patients were randomly grouped into three equal groups of 7 patients. Two groups were used for training and the remaining one group was used for testing, which was repeated three times to ensure testing of each group once. Different data augmentation strategies also were used to increase training samples.

Results: The central distance difference (mm) between the automatic and manual digitization averaged over 21 patients for needles 1 to 7 was 0.69±0.25, 0.68±0.23, 0.57±0.14, 0.63±0.39, 0.61±0.30, 0.62±0.36, and 0.70±0.45, respectively. For individual cases, with an exception of case #2 and case #7, central distance difference between predicted and ground-truth digitization of applicator were all less than 1 mm in remaining 19 test cases. This demonstrates consistency and effectiveness of our proposed method.

Conclusion: A new target activation network-based method was proposed for GYN HDR applicator digitization on CT scans. Preliminary tests found that our method achieved submillimeter accuracy when compared to manual digitization.

Keywords

Segmentation

Taxonomy

TH- Brachytherapy: HDR Brachytherapy

Contact Email

Share: