MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Neural Network Application to Proton Imaging

C Dema1*, U Akgun2 , (1) William Jewel College, Liberty, Missouri, (2) Coe College, Cedar Rapids, IA

Presentations

(Sunday, 7/29/2018) 3:00 PM - 6:00 PM

Room: Exhibit Hall

Purpose: To develop a neural network proton imaging method for a high density scintillating glass detector system. The addition of machine learning, with proper training, aims to shorten the image reconstruction time.

Methods: A compact, glass calorimeter was designed as the gantry-mounted proton imager. The novel high-density scintillating glasses are developed for this detector; require 7 cm thickness to stop 250 MeV protons. The capabilities of the detector were tested with Shepp-Logan phantom in Geant4 simulation package. The straight line and cubic spline path estimations were successfully implemented to reconstruct the 2D images. Generated Geant4 simulation data sets were combined with the 1 degree slices of radon transformation images for training the neural network in MATLAB AI package and Google’s TensorFlow. The geometrical, and density variations were created on the phantom to generate more training sets and verification data for the neural network algorithms.

Results: A possible application of machine learning to proton imaging based on a specific geometry is demonstrated. The image reconstruction times of the neural network algorithms are compared to that of the standard methods based on straight line and cubic spline path estimations. The tumor size and shape prediction capabilities of different algorithms are compared in MATLAB AI Package, and the TensorFlow. As the algorithms are continuously trained with new tumor models their imaging capabilities are shown to improve drastically.

Conclusion: This work shows the potential of using neural network in proton imaging. The proof-of-concept studies have been completed for a compact proton imager, and the machine-learning algorithms are shown to successfully reconstruct the image of the given “unknown� phantom. This approach shortens the image reconstruction time compared to the standard methods. As the algorithm is trained with variety of the phantoms, its capability to recognize the density, and the shape of the tumor improves.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email