Click here to

Session: Adaptive and Biology-guided Radiation Therapy [Return to Session]

Deep Learning-Based Automatic Segmentation Framework for Targets and OARs in Cervical High Dose Rate (HDR) Brachytherapy

R Ni1*, S Kim1, B Haibe-kains1, A Rink1,2, (1) University of Toronto, Toronto, ON, CA, (2) Princess Margaret Hospital, Toronto, ON, CA

Presentations

TU-D1000-IePD-F7-1 (Tuesday, 7/12/2022) 10:00 AM - 10:30 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 7

Purpose: Brachytherapy is an important component in managing cervical cancer and improves survival rates over external beam radiotherapy alone. Manually accurate delineation of organs at risk (OARs) and targets is the most time-consuming aspect of a cervix brachytherapy procedure. To reduce the contouring time, this study aims to develop an automatic segmentation framework in female pelvic magnetic resonance (MR) images using deep learning (DL) techniques.

Methods: A dataset of 48 T2-weighted MR images (acquired axially) of cervix region was built from 13 cervical cancer patients undergoing brachytherapy. These MR images were manually contoured by radiation oncologists during brachytherapy treatment. An end-to-end automatic segmentation framework was developed based on a novel self-adapting U-Net-based method, called nnU-Net ('no-new-Net'). 3D full-resolution U-Net architecture was manually selected to preserve the image quality and the automated configuration was used for preprocessing, network training, and postprocessing. The network was trained on 42 cases from 11 patients for 200 epochs with 5-fold cross-validation and applied to four OARs (bladder, rectum, sigmoid, and small bowel) and high-risk clinical target volume (HR-CTV). The segmentation performance was evaluated by the volumetric Dice Similarity Coefficients (DSC), 95th percentile of Hausdorff distance (HD95), surface DSC, and added path length (APL).

Results: The preliminary results of the automatic segmentation showed that the volumetric DSCs (mean±standard deviation) for the 6 test cases were 0.87±0.03 (bladder), 0.83±0.03 (rectum), 0.64±0.08 (sigmoid), 0.6±0.2 (small bowel), and 0.6±0.1 (HR-CTV). With the novel self-configuration method, our model demonstrated comparable performance with previous studies despite the very limited dataset size.

Conclusion: We have proposed a DL-based automatic segmentation approach for cervical cancer targets and OARs in MR-guided brachytherapy. This method holds great potential to improve contouring efficiency without clinically significant loss of accuracy and thus, improving plan accuracy and decreasing dosimetric uncertainty.

Funding Support, Disclosures, and Conflict of Interest: CIHR #426366; Funding support from TECHNA Institute

Keywords

Brachytherapy, MRI, Segmentation

Taxonomy

TH- Brachytherapy: Imaging for brachytherapy: development and applications

Contact Email

Share: