Personne : Archambault, Louis
En cours de chargement...
Adresse électronique
Date de naissance
Projets de recherche
Structures organisationnelles
Fonction
Nom de famille
Archambault
Prénom
Louis
Affiliation
Université Laval. Département de physique, de génie physique et d'optique
ISNI
ORCID
Identifiant Canadiana
ncf11853180
person.page.name
3 Résultats
Résultats de recherche
Voici les éléments 1 - 3 sur 3
- PublicationAccès libreOn the proper use of structural similarity for the robust evaluation of medical image synthesis models(American Association of Physicists in Medicine by the American Institute of Physics, 2022-02-14) Duchesne, Simon; Gourdeau, Daniel; Archambault, LouisPurpose To propose good practices for using the structural similarity metric (SSIM) and reporting its value. SSIM is one of the most popular image quality metrics in use in the medical image synthesis community because of its alleged superiority over voxel-by-voxel measurements like the average error or the peak signal noise ratio (PSNR). It has seen massive adoption since its introduction, but its limitations are often overlooked. Notably, SSIM is designed to work on a strictly positive intensity scale, which is generally not the case in medical imaging. Common intensity scales such as the Houndsfield units (HU) contain negative numbers, and they can also be introduced by image normalization techniques such as the z-normalization. Methods We created a series of experiments to quantify the impact of negative values in the SSIM computation. Specifically, we trained a three-dimensional (3D) U-Net to synthesize T2-weighted MRI from T1-weighted MRI using the BRATS 2018 dataset. SSIM was computed on the synthetic images with a shifted dynamic range. Next, to evaluate the suitability of SSIM as a loss function on images with negative values, it was used as a loss function to synthesize z-normalized images. Finally, the difference between two-dimensional (2D) SSIM and 3D SSIM was investigated using multiple 2D U-Nets trained on different planes of the images. Results The impact of the misuse of the SSIM was quantified; it was established that it introduces a large downward bias in the computed SSIM. It also introduces a small random error that can change the relative ranking of models. The exact values for this bias and error depend on the quality and the intensity histogram of the synthetic images. Although small, the reported error is significant considering the small SSIM difference between state-of-the-art models. It was shown therefore that SSIM cannot be used as a loss function when images contain negative values due to major errors in the gradient calculation, resulting in under-performing models. 2D SSIM was also found to be overestimated in 2D image synthesis models when computed along the plane of synthesis, due to the discontinuities between slices that is typical of 2D synthesis methods. Conclusion Various types of misuse of the SSIM were identified, and their impact was quantified. Based on the findings, this paper proposes good practices when using SSIM, such as reporting the average over the volume of the image containing tissue and appropriately defining the dynamic range.
- PublicationAccès libreDiscriminative neural network for hero selection in professional Heroes of the Storm and DOTA 2(Institute of Electrical Engineers Inc., 2020-02-07) Gourdeau, Daniel; Archambault, LouisMultiplayer online battle arena games (MOBAs) are one of the most popular types of online games. Annual tournaments draw large online viewership and reward the winning teams with large monetary prizes. Character selection prior to the start of the game (draft) plays a major role in the way the game is played and can give a large advantage to either team. Hence, professional teams try to maximize their winning chances by selecting the optimal team composition to counter their opponents. However, drafting is a complex process that requires deep game knowledge and preparation, which makes it stressful and error-prone. In this paper, we present an automatic drafter system based on the suggestions of a discriminative neural network and evaluate how it performs on the MOBAs Heroes of the Storm and DOTA 2. We propose a method to appropriately exploit very heterogeneous datasets that aggregates data from various versions of the games. Drafter testing on professional games shows that the actual selected hero was present in the top 3 determined by our drafting tool 30.4% of the time for HotS and 17.6% for DOTA 2. The performance obtained by this method exceed all previously reported results.
- PublicationAccès libreAn EPID-based method to determine mechanical deformations in a linear accelerator(American Association of Physicists in Medicine by the American Institute of Physics, 2018-09-22) Gingras, Luc; Leclerc, Ghyslain; Beaulieu, Frédéric; Gourdeau, Daniel; Archambault, LouisPurpose: Medical linear accelerators (linac) are delivering increasingly complex treatments using modern techniques in radiation therapy. Complete and precise mechanical QA of the linac is therefore necessary to ensure that there is no unexpected deviation from the gantry's planned course. However, state-of-the-art EPID-based mechanical QA procedures often neglect some degrees of freedom (DOF) like the in-plane rotations of the gantry and imager or the source movements inside the gantry head. Therefore, the purpose of this work is to characterize a 14 DOF method for the mechanical QA of linacs. This method seeks to measure every mechanical deformation in a linac, including source movements, in addition to relevant clinical parameters like mechanical and radiation isocenters. Methods: A widely available commercial phantom and a custom-made accessory inserted in the linac's interface mount are imaged using the electronic portal imaging device (EPID) at multiple gantry angles. Then, simulated images are generated using the nominal geometry of the linac and digitized models of the phantoms. The nominal geometry used to generate these images can be modified using 14 DOF (3 rigid rotations and 3 translations for the imager and the gantry, and 2 in-plane translations of the source) and any change will modify the simulated image. The set of mechanical deformations that minimizes the differences between the simulated and measured image is found using a genetic algorithm coupled with a gradient-descent optimizer. Phantom mispositioning and gantry angular offset were subsequently calculated and extracted from the results. Simulations of the performances of the method for different levels of noise in the phantom models were performed to calculate the absolute uncertainty of the measured mechanical deformations. The measured source positions and the center of collimation were used to define the beam central axis and calculate the radiation isocenter position and radius. Results: After the simultaneous optimization of the 14 DOF, the average distance between the center of the measured and simulated ball bearings on the imager was 0.086 mm. Over the course of a full counter-clockwise gantry rotation, all mechanical deformations were measured, showing sub-millimeter translations and rotations smaller than 1° along every axis. The average absolute uncertainty of the 14 DOF (1 SD) was 0.15 mm or degree. Phantom positioning errors were determined with more than 0.1 mm precision. Errors introduced in the experimental setup like phantom positioning errors, source movements or gantry angular offsets were all successfully detected by our QA method. The mechanical deformations measured are shown to be reproducible over the course of a few weeks and are not sensitive to the experimental setup. Conclusion: This work presents of new method for an accurate mechanical QA of the linacs. It features a 14 DOF model of the mechanical deformations that is both more complete and precise than other available methods. It has demonstrated sub-millimeter accuracy through simulation and experimentation. Introduced errors were successfully detected with high precision.