D XU^1,2*, T Ma², R Savjani², M Cao², Y Yang², A Kishan², F Scalzo³, K Sheng², (1)Computer Science, University of California, Los Angeles, CA 90095, USA (2) Radiation Oncology, University of California, Los Angeles, CA 90095, USA (3)Computer Science, Pepperdine University, 24255 Pacific Coast Hwy, Los Angeles, CA 90263, USA

WE-B-201-3 (Wednesday, 7/13/2022) 8:30 AM - 9:30 AM [Eastern Time (GMT-4)]

Room 201

Purpose: Accurate delineation of the intra-prostatic urethra is a prerequisite for urinary dose reduction in prostate cancer radiotherapy. However, manual delineation of Foley catheter-free urethra is resource-demanding and inconsistent. This work develops and tests a fully automated pipeline for efficient and consistent urethra segmentation in magnetic resonance-guided radiation therapy (MRgRT).

Methods: Twenty-eight prostate cancer patients who underwent MRgRT were included in this study. In-house 3D HASTE and 3D TSE sequences were used to image the Foley-free urethra on a 0.35T MRgRT system (MRIdian, ViewRay). The segmentation pipeline uses 3D nnU-Net as the backbone network and innovatively combines ground truth and its corresponding radial distance (RD) map during training supervision. In addition, we evaluated the benefit of incorporating a novel LSTM-Conv layer and spatial recurrent layers (RCL) into nnU-Net. A novel slice-by-slice simple exponential smoothing (SEPS) method specifically for 3D tubular structures was used to post-process the segmentation results. Dice’s coefficient and the Hausdorff distance (HD) were used to evaluate the automated segmentation results.

Results: The experimental results show that 3D nnU-Net trained using a combination of Dice, cross-entropy (CE) and radial distance (RD) losses achieved a Dice score of 76.2% in the testing dataset. With further implementation of SEPS, while maintaining equivalent Dice score, Hausdorff distance (HD) and 95% HD were reduced to 2.95 mm and 1.84 mm, respectively. Lastly, LSTM and RCL layers only minimally improved the segmentation precision.

Conclusion: We present the first Foley-free automated urethra segmentation study on 0.35 T MR images generated during MRgRT. The automated segmentation method is built on a data-driven deep learning neural network with novel cost functions and a post-processing step specifically designed for the tubular urethra structure. The segmentation performance is consistent with the need for online and offline urethra dose reduction in prostate radiotherapy.

Funding Support, Disclosures, and Conflict of Interest: funding support from NIH R01CA259008, DOD W81XWH2210044

Keywords

MRI, Computer Vision, Segmentation

Taxonomy

IM- MRI : Machine learning, computer vision

Contact Email