A Generative Adversarial Network (GAN)-based Approach for Automatic Fiducial Marker Detection in MR Radiotherapy Simulation Images

Abstract

To design and evaluate a robust method for the automatic detection of fiducial markers (FMs) in MR simulation images used for prostate cancer radiotherapy planning. MR-only simulation continues to see expanded use due to the benefits of eliminating CT from standard MR/CT radiotherapy planning workflows, including reduced registration uncertainties, cost, and time. Standard x-ray-based IGRT treatments that use FMs for patient alignment pose a challenge for MR-only simulation, where FMs appear as small and often difficult to detect signal voids. In practice, multiple MR sequences with varying echo times and manual interaction are often required to distinguish FMs. Such methods are vulnerable to patient motion and require increased clinician effort. In this work we present an automatic FM detection method requiring a single standard GRE-based MR sequence. 21 prostate cancer patients who each had three FMs implanted prior to CT simulation and MR imaging were retrospectively selected for this study. Contrast enhanced T1 VIBE images were used in this study. Ground truth FM contours were manually defined on the patients’ CT and corresponding MR images. The cohort was divided into five groups, with one group held out for testing and the other four used for training in a five-fold cross validation approach. The MR images and FM contours (converted to labeled binary masks) from the training sets were used to train a Generative Adversarial Network (GAN) with the pix2pix image-to-image translation package. Trained models take an input MR image and output an image with suspected FM pixels labeled. The final three FM contours for each patient were defined as the three largest clustered objects in the labeled volume. Successful FM identification was defined based on the DICE coefficient of the automatically detected and ground truth FM contours. The Fiducial Registration Error (FRE) between the automatically detected and ground truth FM locations were calculated. FMs were correctly identified in 97% of tested cases (Table 1). All implanted FMs were correctly identified in 19/21 of the tested patients. The two incorrect FM identifications occurred with a marker implanted at the prostate/rectal wall interface and urethra. The mean FRE difference in the Anterior-Posterior, Left-Right and Superior-Inferior directions was 0.3 ± 0.7, 0.7 ± 1.4 and 0.8 ± 1.1 mm respectively. We developed a GAN-based approach to automatic FM detection in MR simulation images, and used it to demonstrate the feasibility of detecting FMs with a single standard MR sequence. Future work will involve testing the model on a larger cohort of patients and further investigating potential sources of FM misidentification, including calcifications and proximity to tissue interfaces.Abstract 189; Table 1Group 1 (Patient 1-4)Group 2 (Patient 5-8)Group 3 (Patient 9-12)Group 4 (Patient 13-17)Group 5 (Patient 18-21)% FM Detection accuracy1001001009292ΔFRE Magnitude (mm)1.4 ± 1.90.7 ± 11 ± 1.61.5 ± 1.81.6 ± 2.6 Open table in a new tab