Program Information
Fully Automated Image Analysis for MV Imaging and Treatment Isocenter Coincidence Using a Feed-Forward Neural Network
P Florio*, M Reyhan , Thomas Jefferson Univ Hospital, Philadelphia, PA
Presentations
SU-I-GPD-J-113 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose: To automate image analysis of planar MV imaging and treatment isocenter coincidence for monthly quality assurance
Methods: Monthly quality assurance (QA) of MV planar imaging and treatment coincidence was performed using the Penta-Guide phantom aligned to in-room lasers and portal imaging using an Elekta Agility accelerator. Passing criteria was defined by visual inspection of the contrast region within the phantom and its coincidence with isocenter (tolerance < 1 mm). A pattern recognition two layer feed-forward neural network (NN) with sigmoid hidden and softmax output neurons was trained using scaled conjugate gradient backpropagation (Matlab). The NN inputs include: the distance from the center of the radiation field to the center of the phantom’s contrast region (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 NN was trained, tested, and validated using 21 images from previous monthly QA testing and failed images generated for training the network. Additionally, 100 images were generated (50 passing) and processed by the NN for validation against the ‘gold standard’ visual inspection.
Results: The NN results achieved 100% accuracy compared with the ‘gold-standard’, when given the additional 100 images to categorize (pass/fail). A paired t-test was performed and demonstrated no significant difference. Sensitivity was 100% and specificity also 100%. The positive predictive value was 1 and the negative predictive values was 1. Linear regression analysis produced and equation of: Gold-Standard=1*Neural-Network, with a Pearson correlation coefficient of 1. The NN algorithm demonstrated sensitivity to inaccurate setup, changes in image and contrast resolution.
Conclusion: A novel image processing algorithm for MV imaging and treatment isocenter coincidence was validated and demonstrated excellent agreement with the ‘gold-standard’. This algorithm helps remove the subjective nature of imaging quality assurance.
Contact Email: