Loading Events

« All Events

  • This event has passed.

ECE MS Thesis Defense: Matin Raayai Ardakani

April 21, 2021 @ 10:00 am - 11:00 am

MS Thesis Defense: A Framework for Denoising Two and Three-dimensional Monte CarloPhoton Transport Simulations Using Convolutional Neural Networks

Matin Raayai Ardakani

Location: Zoom Link

Abstract: The Monte Carlo (MC) method is considered to be the gold standard for modeling light propagation inside turbid media, proving superior to other Radiative Transfer Equation (RTE) solvers relying on variational principles. However, like most MC-based algorithm, a large number of independently launched photons is needed for converging to the correct result and combating its inherent stochastic noise, yielding longer computation times, even when accelerated on GraphicProcessing Units (GPUs).
To remove this noise from the output without increasing the number of photons used for simulation, modified versions of commonly used filters for image and volumetric data based on non-local self similarity has been used in the past. Current state-of-the-art denoising approaches rely on Convolutional Neural Networks (CNN) to remove spatially variant noise, but the high dynamic range of MC simulations has hindered their adaptation to remove MC noise.
In this thesis, we address this problem by presenting a supervised framework for using CNNs to denoise MC simulations. First, a dataset is created with each entry comprising of a unique configuration simulated with different numbers of photons. The simulation configurations are generated using a simple generative model that introduces objects with both smooth and sharp edges into the volume. By selecting the group of fluence maps simulated with the maximum number of photons in the dataset as labels, we train a range of CNN-based models to learn the underlying mapping between noisy and clean images. The CNN input is converted to log scale and normalized to reduce the high dynamic range, and converted back after inference. The trained CNNs are then shown to have better performance compared to using an Adaptive Non-local Means filter, in terms of mean square error (MSE), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR) in the image domain.
Finally, we purpose our own architecture that combines DnCNN and UNet, a strategy that can learn both local and global residual noise maps, achieving state-of-the-art performance compared to existing CNN methods. Future avenues of research and challenges for denoising 3D simulations are also discussed.

Details

Date:
April 21, 2021
Time:
10:00 am - 11:00 am
Website:
https://northeastern.zoom.us/j/98095326959#success

Other

Department
Electrical and Computer Engineering
Topics
MS/PhD Thesis Defense