Click here to


Are you sure ?

Yes, do it No, cancel

Deep Learning Techniques in Microdosimetry: Using Conditional Generative Adversarial Networks to Predict Energy Deposition On Cellular Length Scales

I Mansour*, R Thomson, Carleton Univ, Ottawa, ON


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Studies of cellular radiation response traditionally use experimental and Monte Carlo (MC) methods, and both can present diverse challenges. Focusing on computational aspects, we investigate the generation of realistic specific energy distributions on cellular length scales using a conditional generative adversarial network (CGAN) trained using MC-generated datasets.

Methods: A CGAN is trained using MC-generated distributions of specific energy (energy imparted per unit mass) scored in microscopic (1-11 micron) water voxels irradiated by photon sources (20-150 keV). Different dose levels (mean specific energies) are considered. The CGAN-generated distributions are assessed based on comparisons with MC data, considering the generated mean, standard deviation, microdosimetric spread (quotient of standard deviation and mean) and number of voxels receiving no energy.

Results: The CGAN is capable of producing any state within the training data domain, and relative errors in comparison with MC specific energy distributions depend on dose level. For example, considering the relatively low dose of 8 mGy (for which the microdosimetric spread is considerable at >100% for all targets, source energies), the mean relative errors over all target sizes and source energies are 9% (specific energy mean), 14% (standard deviation), and 20% (number of targets receiving no energy); these decrease with increasing dose, e.g. with corresponding mean relative errors of 4%, 6%, and 14%, respectively, at 20 mGy. Once trained, the CGAN can generate specific energy distributions much faster than MC: on average, 3.4 x 104 times faster the MC.

Conclusion: Trained CGAN successfully generates specific energy distributions in good agreement with MC data, capturing variations with source energy, voxel size, and dose level. CGANs offer a means of precomputing microscopically relevant situations for later use, providing an efficient alternative to computationally-intensive MC simulations. Ongoing work is developing CGANs for computing energy deposition in populations of cells with representative non-water elemental compositions.

Funding Support, Disclosures, and Conflict of Interest: Funding: Natural Sciences and Engineering Research Council (NSERC) of Canada Canada Research Chairs Ontario Ministry of Research and Innovation COI: The authors have no conflicts of interest to disclose.


Microdosimetry, Monte Carlo, Machine Learning


IM/TH- Radiation Transport: Microdosimetry (computational)

Contact Email