Program Information
Fully Automated Image Analysis for Daily Winston-Lutz Testing Using a Feed-Forward Neural Network
P Florio*, H Liu , M Reyhan , Thomas Jefferson Univ Hospital, Philadelphia, PA
Presentations
SU-I-GPD-J-23 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose: To validate the accuracy of a feed-forward neural network approach for daily imaging based Winston-Lutz QA
Methods: Winston-Lutz quality assurance (QA) utilizing MV imaging was performed using a ball-bearing (BB) phantom (0.5cm diameter) and portal imaging with a Varian TruebeamSTx to determine isocentric accuracy. Images were acquired at various gantry angles. ‘Gold-standard’ passing criteria was defined by Radiologic Imaging Technology(RIT) software analysis with a tolerance < 1 mm difference between the centers of the BB and the radiation field. A pattern recognition neural network was trained using a two layer feed-forward neural network with sigmoid hidden and softmax output neurons using scaled conjugate gradient backpropagation on monthly MV imaging and treatment isocenter coincidence tests using a Penta-Guide phantom and Elekta Agility. The network was trained, tested, and validated using 21 images from monthly QA testing. The network inputs include: the distance from the center of the radiation field to the center of the ball bearing (determined by morphologic image processing), the maximum value from 2D-normalized cross-correlation relative to a baseline image, the matrix location of the maximum value from 2D-normalized cross-correlation, and the structural similarity index. The morphologic image processing algorithm was updated to account for the BB phantom. The 300 Winston-Lutz images (11 failing) were retrospectively processed by the neural network for statistical validation against the ‘gold standard’ results.
Results: The neural network based algorithm achieved 99.7% (299/300) classification accuracy, misclassifying 1 passing image as a fail. The sensitivity was 0.9965, specificity was 1, correct classification rate was 0.9967, positive predictive power was 1, negative predictive power was 0.9167, and kappa was 0.9548.
Conclusion: A novel image processing algorithm based on neural network analysis for Winston-Lutz based isocentric QA was validated and demonstrated excellent agreement with the current ‘gold-standard’. It's a promising tool for automated daily QA analysis.
Contact Email: