This work aims at developing a generic virtual source model (VSM) preserving all existing correlations between variables stored in a Monte Carlo pre-computed phase space (PS) file, for dose calculation and high-resolution portal image prediction. The reference PS file was calculated using the PENELOPE code, after the flattening filter (FF) of an Elekta Synergy 6 MV photon beam. Each particle was represented in a mobile coordinate system by its radial position (r s ) in the PS plane, its energy (E), and its polar and azimuthal angles (φ d and θ d ), describing the particle deviation compared to its initial direction after bremsstrahlung, and the deviation orientation. Three sub-sources were created by sorting out particles according to their last interaction location (target, primary collimator or FF). For each sub-source, 4D correlated-histograms were built by storing E, r s , φ d and θ d values. Five different adaptive binning schemes were studied to construct 4D histograms of the VSMs, to ensure histogram efficient handling as well as an accurate reproduction of E, r s , φ d and θ d distribution details. The five resulting VSMs were then implemented in PENELOPE. Their accuracy was first assessed in the PS plane, by comparing E, r s , φ d and θ d distributions with those obtained from the reference PS file. Second, dose distributions computed in water, using the VSMs and the reference PS file located below the FF, and also after collimation in both water and heterogeneous phantom, were compared using a 1.5%-0 mm and a 2%-0 mm global gamma index, respectively. Finally, portal images were calculated without and with phantoms in the beam. The model was then evaluated using a 1%-0 mm global gamma index. Performance of a mono-source VSM was also investigated and led, as with the multi-source model, to excellent results when combined with an adaptive binning scheme.
Low-field magnetic resonance (MR) images suffer from inherent low Signal-to-noise ratio (SNR) compared to images acquired using high-field MRI scanners. Denoising these images could help and improve further processing, such as image segmentation. In this paper a Convolutional Neural Network AutoEncoder was designed with a dedicated loss function for noise reduction. A transfer learning approach was employed in which high-field high-SNR MR images, served as targets for learning from their noise-added counterparts. Evaluation of network performances was measured on both noisy high-field and low-field MR images that had not been included in the learning step. The proposed method outperformed major denoising methods applied to MR images. SNR improvements were quantified on low-field MR images.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.