Loading...Loading...

Deep learning approach for denoising Monte Carlo dose distribution in proton therapy

Asensi Madrigal, Jorge

(2018)

Details

Supervisors

Lee, John Aldo ; Souris, Kevin ; Sterpin, Edmond ; Labarbe, Rudi

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil biomédical

Abstract

Radiation therapy planning requires to simulate the dose distribution on a CT patient image. It is used an algorithm based on Monte Carlo to generate that simulation, but this algorithm produces some noise that need to be removed. Convolutional Neural Networks (CNN) have improved the state-of-the-art in the recent years by recognizing hierarchical features on an image. The purpose of the current work is to build a Neural Network that take a 3D Monte Carlo Dose Distribution as an input and denoise it through the different layers it includes, to use it in hospital practice. The quality of Monte Carlo generated images depends on the number of particles employed, consequently, improving the quality of the images involves an exponential increase of the computing time. Simulations generated with 1e9 particles could be considered as free-noisy because the residual noise they have does not compromise the clinical application. We filtered distributions generated with 1e7 and 1e6 particles, what result in one minute and 10 seconds of computing, respectively. Both networks architectures are U-Net, commonly used in the segmentation task, both of them exceed the state-of-the-art, achieving a signal-to-noise ratio of 73.03 and 35.69 respectively, and they spend 45 seconds around on filtering the whole 3D-image.