(Saturday, 4/4/2020) [Mountain Time (GMT-6)]
Purpose: To create and train a Convolutional Neural Network (CNN) capable of detecting dead detector elements (DDEs) in diagnostic digital detectors.
Methods: A CNN was created by utilizing the open source software Tensorflow and Keras. The CNN was coded and is customizable to include a non-model generated feature map incorporated to help identify DDEs. The non-model generated map is determined by incorporating properties of the Noise Power Spectrum (NPS). Any averaging techniques performed by the vendor in the presence of a DDE should result in changes to NPS factors.Flat field images were acquired from an on board imaging (OBI) system on a LINAC, with various kVp and mAs combinations. The full images were then broken into 28x28 pixel sub units which serve as samples for the CNN. The dead pixel map was utilized for the specific OBI system and served as the ground truth for samples in the network. The CNN was tasked with categorizing each 28x28 sample by identifying whether the sample contained a singular DDE, two DDEs, up to 5% DDEs or between 5% and 10% DDEs.
Results: A CNN model including a non-model generated feature map has been created, trained and tested. Results show 95% training accuracy and 95% verification accuracy. Investigation of results are compared to ground truth data. Further investigations into overtraining and revised training parameters is discussed.
Conclusion: A CNN artificial intelligence network may provide medical physicists with a tool to test the integrity of their digital detection system without a reliance on the vendor. By training the CNN on samples with DDEs, it may identify DDEs from clinical diagnostic systems.
ePosters
Contact Email
