arXiv subjects

Brian W. Pogue

Publications and source records attributed to Brian W. Pogue.

20 records · Page 2Linked to original sources

arXiv2020

Monte Carlo modeling photon-tissue interaction using on-demand cloud infrastructure

Purpose: This work advances a Monte Carlo (MC) method to combine ionizing radiation physics with optical physics, in a manner which was implicitly designed for deployment with the most widely accessible parallelization and portability possible. Methods: The current work updates a previously developed optical propagation plugin for GEANT4 architecture for medically oriented simulations (GAMOS). Both virtual-machine (VM) and container based instances were validated using previously published scripts, and improvements in execution time using parallel simulations are demonstrated. A method to programmatically deploy multiple containers to achieve parallel execution using an on-demand cloud-based infrastructure is presented. Results: A container-based GAMOS deployment is demonstrated using a multi-layer tissue model and both optical and X-ray source inputs. As an example, the model was split into 154 simulations which were run simultaneously on 64 separate containers across 4 servers. Conclusions: The container-based model provides the ability to execute parallel simulations of applications which are not inherently thread-safe or GPU-optimized. In the current demonstration, this reduced the time by at most 97% compared to sequential execution. The code and examples are available through an interactive online interface through links at: https://sites.dartmouth.edu/optmed/research-projects/monte-carlo-software/

physics.med-ph↗

arXiv2019

Coloring the Black Box: Visualizing neural network behavior with a self-introspective model

The following work presents how autoencoding all the possible hidden activations of a network for a given problem can provide insight about its structure, behavior, and vulnerabilities. The method, termed self-introspection, can show that a trained model showcases similar activation patterns (albeit randomly distributed due to initialization) when shown data belonging to the same category, and classification errors occur in fringe areas where the activations are not as clearly defined, suggesting some form of random, slowly varying, implicit encoding occurring within deep networks, that can be observed with this representation. Additionally, obtaining a low-dimensional representation of all the activations allows for (1) real-time model evaluation in the context of a multiclass classification problem, (2) the rearrangement of all hidden layers by their relevance in obtaining a specific output, and (3) the obtainment of a framework where studying possible counter-measures to noise and adversarial attacks is possible. Self-introspection can show how damaged input data can modify the hidden activations, producing an erroneous response. A few illustrative are implemented for feedforward and convolutional models and the MNIST and CIFAR-10 datasets, showcasing its capabilities as a model evaluation framework.

cs.LG↗