3347 - Prostate Auto-Segmentation Model for MRI with or without Hydrogel Spacer
Presenter(s)
Y. Song1,2, L. Nguyen2,3, A. Dornisch2,4, M. Baxter2, D. Do2, I. Li3, T. Barrett5, R. T. Dess6, M. Harisinghani7, S. C. Kamran8, M. A. Liss9, D. Margolis10, E. P. Weinberg11, S. A. Woolen12, A. M. Dale13,14, and T. M. Seibert2,15; 1Department of Electrical and Computer Engineering, University of California San Diego, La Jolla, CA, 2Department of Radiation Medicine and Applied Sciences, University of California San Diego, La Jolla, CA, 3University of California San Diego School of Medicine, La Jolla, CA, 4Center for Health Education and Research, University of California, San Diego, La Jolla, CA, 5Department of Radiology, University of Cambridge, Cambridge, United Kingdom, 6Department of Radiation Oncology, University of Michigan, Ann Arbor, MI, 7Department of Radiology, Massachusetts General Hospital, Boston, MA, 8Department of Radiation Oncology, Massachusetts General Hospital, Harvard Medical School, Boston, MA, 9Department of Urology, University of California San Diego, La Jolla, CA, 10Department of Radiology, Cornell University, Ithaca, NY, 11Department of Clinical Imaging Sciences, University of Rochester Medical Center, Rochester, NY, 12Department of Radiology and Biomedical Imaging, University of California San Francisco, San Francisco, CA, USA, San Francisco, CA, 13Department of Radiology, University of California San Diego, La Jolla, CA, 14Department of Neurosciences, University of California San Diego, La Jolla, CA, 15Department of Bioengineering, University of California San Diego, La Jolla, CA
Purpose/Objective(s): MRI is superior to CT for delineation of the prostate. MRI also makes it possible to contour the urethra and primary prostate cancer tumors. Automated segmentation tools for the prostate could improve clinical efficiency, reproducibility, and possibly even accuracy compared to CT-only treatment planning. However, many patients have a hydrogel spacer placed to reduce rectal toxicity, and prostate auto-segmentation accuracy has not been established for MRI that may or may not include a spacer. We developed and evaluated a prostate segmentation model that incorporates spacer segmentation to improve accuracy.
Materials/Methods: This multi-institutional retrospective study compared two models based on the nnUNet architecture: Model 1 (prostate-only auto-segmentation) versus Model 2 (including two sub-models: prostate-only auto-segmentation from Model 1 and a spacer-only auto-segmentation model to correct spacer-related errors in the initial prostate segmentation). Training data included axial T2-weighted MRI from 659 patients (41 had a spacer). Prostate and spacer structures had been contoured previously per clinical routine. Model performance was assessed using Dice scores and volume difference against manual prostate contours in an independent dataset of 84 T2-weighted MRI cases from 6 institutions (2020-2024; 16 had a spacer). Statistical significance was assessed using t-tests (two-sided alpha=0.05).
Results: Both Model 1 and Model 2 performed well for prostate segmentation in testing dataset cases without a spacer (n=68; Table 1). When a spacer was present (n=16), Model 1 often incorrectly included some of the spacer as prostate, leading to lower Dice and larger volume difference. Model 2 segmented the spacer, when present, and adjusted prostate contours accordingly, yielding slightly higher Dice (p=0.01) and significantly smaller volume difference (p=0.03) than Model 1.
Conclusion: Prostate auto-segmentation on MRI can include spacer in the prostate contours. Auto-segmentation of the spacer can correct these errors in patients with a spacer while preserving performance in spacer-free patients. This dual-stage approach could lead to a robust prostate auto-segmentation solution for patients with or without the presence of spacer. Model 1 is a prostate-only auto-segmentation model. Model 2 is a dual prostate-spacer segmentation model with correction of spacer-related errors. All results in this table are for independent testing data not used in model training.
Abstract 3347 - Table 1: Dice scores and absolute Volume Difference (%) of prostate segmentation in patients with or without a spacer| Model | Dataset | Dice | Volume difference (%) |
| Model | Dataset | Mean ± Standard Deviation | Mean ± Standard Deviation |
| Model 1 | No Spacer (n=68) | 0.93 ± 0.02 | 6.3 ± 5.1 |
| Model 2 | No Spacer (n=68) | 0.93 ± 0.02 | 6.4 ± 5.1 |
| Model 1 | Spacer (n=16) | 0.87 ± 0.06 | 14.0 ± 15.1 |
| Model 2 | Spacer (n=16) | 0.89 ± 0.03 | 7.3 ± 6.9 |