MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Denoising of Monte Carlo Dose Distributions Using UNet

U Javaid1*, K Souris1 , D Dasnoy1 , A Barragan Montero1 , S Huang2 , J Lee1 , (1) UCLouvain, Belgium,(2) Memorial Sloan Kettering Cancer Center, New York, NY

Presentations

(Sunday, 7/14/2019) 2:00 PM - 3:00 PM

Room: Stars at Night Ballroom 1

Purpose: Monte Carlo (MC) algorithms offer accurate dose calculation by simulating physical interactions of many particles through the patient geometry. However, given their random nature, the resulting dose distributions are affected by statistical noise, which renders difficult their interpretation. This noise can be reduced by simulating a huge number of particles, implying a trade-off between computation time and noise. Previous work on dose denoising is based on smoothening the distributions. In this work, we address the issue using UNet – an encoder-decoder fully convolutional neural network.

Methods: We propose a UNet model that has three down-sampling layers for denoising MC dose distributions. The mean square error is used as loss function to train the model. The network was trained from 31 proton therapy patients (brain, H&N, liver, lungs, prostate) using data augmentation techniques. The open source MCsquare code was used to perform 3 simulations per patient with 1E6 particles. The particle number was increased to 1E9 to compute the reference doses, considered as free of noise.

Results: After training, the model was tested on 4 new patients. MC computation times were 10s and 100min for 1E6 and 1E9 particles, respectively. On average, the D95 calculated from the noisy distributions (1E6 particles) were underestimated by 11.4% (up to 17%) in comparison to the reference (1E9 particles). Using our model, dose maps were denoised in only 7s and reduced the bias to less than 1%. The Peak signal-to-noise ratio (PSNR) was improved by a factor of 11.58dB.

Conclusion: We propose a fast and fully automated UNet-based framework for denoising MC dose distributions. It offers good generalization ability as it involves no pre-processing and can be trained on any tumor site. It provides comparable dose-volume histograms (DVH) to the much longer MC simulation involving 1E9 particles, enabling a significant reduction in computational time.

Funding Support, Disclosures, and Conflict of Interest: Kevin Souris is supported by a research grant from Ion Beam Application (IBA s.a., Louvain-la-Neuve, Belgium).

Keywords

Monte Carlo, Noise Reduction

Taxonomy

TH- External Beam- Particle therapy: Proton therapy - computational dosimetry-Monte Carlo

Contact Email