Room: Room 205
Purpose: To quantify the improvement in fiducial marker detection and position reconstruction utilizing an experimental high-efficiency cadmium tungstate (CWO) detector and a novel convolutional neural network (CNN) based approach.
Methods: An experimental CWO detector, with 20x the detective quantum efficiency (DQE) of a standard gadolinium oxysulfide (GOS) portal imager, was used to collect projection data from a RANDO anthropomorphic phantom implanted with fiducial markers. 20x20 pixel images containing fiducial markers were extracted from the projections. These data were used as a training set for a CNN-based image classifier comprised of two convolutional layers having three 3x3 pixel optimized convolution filters. Images were categorized into two classes - those containing fiducials and a random sampling of regions without fiducials. CNN output was convolved with the kernel of the training-set to produce a well-segmented heat map indicative of fiducial location. Marker positions and ray-projection drove a gradient-based optimization to determine the most likely position of the rigid marker structure. Positions for a static and random couch trajectory (created in developer-mode XML) were reconstructed using this method. Portal dose images from a VMAT prostate delivery were assessed for average offset and rotation for both the GOS and CWO detectors.
Results: The CWO detector provided superior quality in all investigated metrics. CNNs trained using the CWO fiducial images had significantly improved efficiency, lower noise and provided for more accurate and consistent seed position reconstruction. Specifically, areas of high attenuation proved more challenging for the lower efficiency GOS detector and its associated CNN. The analysis pipeline from image to seed position was achieved at a rate of 2Hz - parallelizable to a rate suitable for real-time MV imaging.
Conclusion: A novel technique for fiducial marker position reconstruction was demonstrated indicating the CWO detector to be an enabling technology for this purpose.
Funding Support, Disclosures, and Conflict of Interest: JSL is an employee of Varian Medical Systems who partially funded this work.