Click here to


Are you sure ?

Yes, do it No, cancel

Automatic Fiducial Marker Detection in Prostate Cancer MR Images Using Generative Adversarial Networks (GANs)

K Singhrao1*, J Fu1 , A Kishan1 , J Lewis1 , (1) Dept. of Radiation Oncology, UCLA, Los Angeles, CA,


(Thursday, 7/18/2019) 10:00 AM - 12:00 PM

Room: 225BCD

Purpose: Accurate fiducial marker (FM) localization in MR-only simulation images is required to avoid systematic errors in marker-based patient positioning. FMs appear as small signal voids in MR images, and are often difficult to discern. Existing clinical solutions require multiple MR sequences and/or manual interaction and specialized expertise. Here, we develop a robust single-sequence method for automatic fiducial marker detection in prostate cancer MR-only simulation images.

Methods: A two-model approach was used to first automatically define the prostate search region (PSR), and then localize FMs within the PSR. CT and contrast-enhanced T1-VIBE MR images from 21 patients with three implanted FMs were included in this study. The PSR-definition model was trained using MR images with physician-drawn contours expanded by 1cm. Ground-truth FM contours were defined with assistance from registered CT images, and reviewed by two experts. FM contours, MR images, and the corresponding PSRs were used to train the FM-localization model. Training for both models was performed using the pix2pix Generative Adversarial Network (GAN) image-to-image translation package. Model testing was performed with five-fold cross validation. The Fiducial Registration Error (FRE) between the automatically detected and ground-truth FM locations were calculated. The DICE coefficient between the PSR and the expanded physician-defined prostate contours was also calculated.

Results: FMs were correctly identified in 97% of tested cases. The mean FRE difference in the Anterior-Posterior, Left-Right and Superior-Inferior directions was 0.3 ± 0.7 mm, 0.7 ± 1.4 mm and 0.8 ± 1.1 mm respectively. The mean DICE coefficient between automatically defined PSRs and expanded physician-defined contours was 0.70 ± 0.07.

Conclusion: The presented GAN-based approach to automatic FM detection in MR images was able to accurately identify FMs in nearly all test cases using a single T1-VIBE MR image. Errors in detected marker position were on the order of 1 mm.

Funding Support, Disclosures, and Conflict of Interest: Varian Master Research Agreement


Not Applicable / None Entered.


IM/TH- Image Analysis (Single modality or Multi-modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)

Contact Email