Related Experiment Video
Updated: Jan 6, 2026
05:56
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
3.1K
Mitigating inherent noise in Monte Carlo dose distributions using dilated U-Net.
Umair Javaid1,2, Kevin Souris2, Damien Dasnoy1
1ICTEAM, UCLouvain, Louvain-la-Neuve, 1348, Belgium.
Medical Physics
|October 11, 2019
Summary
This study introduces a novel deep learning approach using a dilated U-Net to effectively denoise Monte Carlo (MC) dose distributions in proton therapy. The method significantly reduces noise and computation time while maintaining dose accuracy for clinical decisions.
Area of Science:
- Medical Physics
- Radiotherapy
- Artificial Intelligence
Background:
- Monte Carlo (MC) simulations provide accurate dose calculations in radiotherapy but suffer from statistical noise.
- This inherent noise in MC dose distributions can hinder reliable clinical decision-making.
- Reducing noise often requires extensive computation time due to the need for a large number of simulated particles.
Purpose of the Study:
- To develop and evaluate a fully convolutional neural network, specifically a dilated U-Net, for automated denoising of MC dose distributions.
- To mitigate the trade-off between computation time and noise level in MC-based dose maps.
- To enable fast and reliable dose distribution analysis in proton therapy.
Main Methods:
- A dilated U-Net architecture was employed as an encoder-decoder fully convolutional neural network.
- The model was trained using Mean Squared Error (MSE) loss in 2D and 2.5D settings on proton therapy MC dose distributions from 35 patients across various tumor sites.
- Input dose distributions simulated with a lower particle count were denoised to approximate those simulated with a higher particle count.
Main Results:
- The trained U-Net model successfully denoised MC dose maps, recovering dose values comparable to high-particle count simulations.
- Average RMSE between reference and denoised dose maps was significantly reduced (1.25 Gy) compared to the noisy input (16.96 Gy).
- The model achieved an 18.06 dB improvement in image signal-to-noise ratio (ISNR) and reduced inference time to under 10 seconds from over 100 minutes for MC simulation.
Conclusions:
- An end-to-end fully convolutional network effectively denoises Monte Carlo dose distributions.
- The proposed network achieves comparable qualitative and quantitative results to high-particle count MC simulations.
- This deep learning approach offers a substantial reduction in computation time for generating high-quality dose distributions.
