Assessing the Ability of Generative Adversarial Networks to Learn Canonical Medical Image Statistics
Not provided.
SEARCH · Search NASA
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
Not provided.
AI components (e.g., Deep Neural Networks) are increasingly used in unmanned Aerospace systems for safety-relevant applications. Rigorous Verification and Validation methods for such components are still in their infancy and thus, monitoring of the AI's behavior during runtime is essential. In this paper, we will present a runtime-monitoring architecture, which combines the advanced statistical analysis framework SYSAI (System Analysis using Statistical AI) with temporal and probabilistic runtime monitoring carried out by R2U2 (Realizable, Responsive, and Unobtrusive Unit). Learned statistical models of complex systems with AI components are produced by the SYSAI framework and provide detailed information to enable the R2U2 runtime monitor to efficiently perform advanced safety and performance checks in nominal and off-nominal conditions. We will present initial results of our tool set and architecture on a case study, a DNN-based autonomous centerline tracking system (ACT).
Purpose: There is tremendous excitement in radiotherapy about applying data-driven methods to develop personalized clinical decisions for real-time response-based adaptation. However, classical statistical learning methods lack in terms of efficiency and ability to predict outcomes under conditions of uncertainty and incomplete information. Therefore, we are investigating physics-inspired machine learning approaches by utilizing quantum principles for developing a robust framework to dynamically adapt treatments to individual patient’s characteristics and optimize outcomes. Methods: We studied 88 liver SBRT patients with 35 on non-adaptive and 53 on adaptive protocols. Adaptation was based on liver function using a split-course of 3+2 fractions with a month break. The radiotherapy environment was modeled as a Markov decision process (MDP) of baseline and one month into treatment states. The patient environment was modeled by a 5-variable state represented by patient’s clinical and dosimetric covariates. For comparison of classical and quantum learning methods, decision-making to adapt at one month was considered. The MDP objective was defined by the complication-free tumor control (P{sup +}=TCPx(1-NTCP)). A simple regression model represented state-action mapping. Single bit in classical MDP and a qubit of 2-superimposed states in quantum MDP represented the decision actions. Classical decision selection was done using reinforcement Q-learning and quantum searching was performed using Grover’s algorithm, which applies uniform superposition over possible states and yields quadratic speed-up. Results: Classical/quantum MDPs suggested adaptation (probability amplitude ≥0.5) 79% of the time for splitcourses and 100% for continuous-courses. However, the classical MDP had an average adaptation probability of 0.5±0.22 while the quantum algorithm reached 0.76±0.28. In cases where adaptation failed, classical MDP yielded 0.31±0.26 average amplitude while the quantum approach averaged a more optimistic 0.57±0.4, but with high phase fluctuations. Conclusion: Our results demonstrate that quantum machine learning approaches provide a feasible and promising framework for real-time and sequential clinical decision-making in adaptive radiotherapy.
Advanced materials with improved properties have the potential to fuel future technological advancements. However, identification and discovery of these optimal materials for a specific application is a non-trivial task, because of the vastness of the chemical search space with enormous compositional and configurational degrees of freedom. Materials informatics provides an efficient approach toward rational design of new materials, via learning from known data to make decisions on new and previously unexplored compounds in an accelerated manner. Here, we demonstrate the power and utility of such statistical learning (or machine learning, henceforth referred to as ML) via building a support vector machine (SVM) based classifier that uses elemental features (or descriptors) to predict the formability of a given ABX 3 halide composition (where A and B represent monovalent and divalent cations, respectively, and X is F, Cl, Br, or I anion) in the perovskite crystal structure. The classification model is built by learning from a dataset of 185 experimentally known ABX 3 compounds. After exploring a wide range of features, we identify ionic radii, tolerance factor, and octahedral factor to be the most important factors for the classification, suggesting that steric and geometric packing effects govern the stability of these halides. As a result, the trained and validated models then predict, with a high degree of confidence, several novel ABX 3 compositions with perovskite crystal structure.
Pulse shape discrimination (PSD) is the task of classifying electronic pulse shapes for different particle types such as gamma rays and fast neutrons interacting in scintillators and read out by photo sensitive detectors. This field has been limited in its adoption of techniques found in the statistical learning community. Methods initially employed in the 1960s for analog electronic circuitry persist in the current PSD literature describing operations performed on digitized pulses, which are amenable to statistical rigor. Despite vast amounts of data collected at low energy levels, traditional PSD methods are unable to discriminate particles below a certain threshold. As such, in this work, Gaussian mixture models (GMMs) are used as a clustering technique for fast neutron detection in the absence of labeled data. GMMs yield improvements spanning the energy spectrum in a desirably efficient, unsupervised fashion. An extension, the Dirichlet Process GMM, provides further flexibility and classification improvement.
Biological processes such as processive enzyme turnover and intracellular cargo tracking involve the dynamic motion of a small "article" along a curvilinear biopolymer track. To understand these processes that occur across multiple length and time scales, one must acquire both the trajectory of the particle and the position of the track along which it moves, possibly by combining high-resolution single-particle tracking with conventional microscopy. Yet, usually there is a significant resolution mismatch between these modalities: while the tracked particle is localized with a precision of 10 nm, the image of the surroundings is limited by optical difraction, with 200 nm lateral and 500 nm axial resolutions. Compared to the particle's trajectory, the surrounding curvilinear structure appears as a blurred and noisy image. This disparity in the spatial resolutions of the particle trajectory and the surrounding curvilinear structure image makes data reconstruction, as well as interpretation, particularly challenging. Analysis is further complicated when the curvilinear structures are oriented arbitrarily in 3D space. Here, we present a prior-apprised unsupervised learning (PAUL) approach to extract information from 3D images where the underlying features resemble a curved line such as a filament or microtubule. This three-stage framework starts with a Hessian-based feature enhancement, which is followed by feature registration, where local line segments are detected on repetitively sampled subimage tiles. In the final stage, statistical learning, segments are clustered based on their geometric relationships. Principal curves are then approximated from each segment group via statistical tools including principal component analysis, bootstrap and kernel transformation. This procedure is characterized on simulated images, where sub-voxel medium deviations from true curves have been achieved. The 3D PAUL approach has also been implemented for successful line localization in experimental 3D images of gold nanowires obtained using a multifocal microscope. Lastly, this work not only bridges the resolution gap between two microscopy modalities, but also allows us to conduct 3D super line-localization imaging experiments, without using super-resolution techniques.
The ability to make rapid and accurate predictions on bandgaps of double perovskites is of much practical interest for a range of applications. While quantum mechanical computations for high-fidelity bandgaps are enormously computation-time intensive and thus impractical in high throughput studies, informatics-based statistical learning approaches can be a promising alternative. Here we demonstrate a systematic feature-engineering approach and a robust learning framework for efficient and accurate predictions of electronic bandgaps of double perovskites. After evaluating a set of more than 1.2 million features, we identify lowest occupied Kohn-Sham levels and elemental electronegativities of the constituent atomic species as the most crucial and relevant predictors. As a result, the developed models are validated and tested using the best practices of data science and further analyzed to rationalize their prediction performance.
The problem of noninstrusive load monitoring (NILM) is usually formulated as a single-channel blind source separation task, whose successful solution enable fast and convenient load identification and energy disaggregation. When applied at test time, NILM algorithms aim to identify the operating characteristics of individual appliances from an aggregate power measurement of the entire house. Recent advances in deep learning gave rise to many methods that mostly focus on learning a direct mapping from aggregate measurement to individual appliance power. However, these methods are not only computationally expensive, but they often suffer from overfitting and do not generalize very well. In this article, we propose a novel NILM method that leverages advances in statistical learning that have not been properly applied in this domain before. The proposed method consists of three stages: first, a Bayesian nonparametric learning-based approach for appliance state extraction; second, synthetic minority oversampling technique for data augmentation and mitigating the heavy imbalance in switching events; and third, appliance-specific lightweight long short-term memory networks for status classification for each appliance. Here, we adopt a “differential” input (the difference before and after the switching event) to reduce the complexity of network training and make the proposed method robust to multiappliance switching events. Experiments are conducted to demonstrate the effectiveness of the proposed method, achieving superior performance when compared to recent methods. An ablation study is conducted to demonstrate the effectiveness of each module of our method. Finally, we investigate the quality of generated synthetic samples.
Molecular dynamics (MD) simulation of complex chemistry typically involves thousands of atoms propagating over millions of time steps, generating a wealth of data. Traditionally these data are used to calculate some aggregate properties of the system and then discarded, but we propose that these data can be reused to study related chemical systems. Using approximate chemical kinetic models and methods from statistical learning, we study hydrocarbon chemistries under extreme thermodynamic conditions. We discover that a single MD simulation can contain sufficient information about reactions and rates to predict the dynamics of related yet different chemical systems using kinetic Monte Carlo (KMC) simulation. Our learned KMC models identify thousands of reactions and run four orders of magnitude faster than MD. In conclusion, the transferability of these models suggests that we can viably reuse data from existing MD simulations to accelerate future simulation studies and reduce the number of new MD simulations required.
Automated segmentation of electron microscopy (EM) images is a challenging problem. In this paper, we present a novel method that utilizes a hierarchical structure and boundary classification for 2D neuron segmentation. With a membrane detection probability map, a watershed merge tree is built for the representation of hierarchical region merging from the watershed algorithm. A boundary classifier is learned with non-local image features to predict each potential merge in the tree, upon which merge decisions are made with consistency constraints in the sense of optimization to acquire the final segmentation. Independent of classifiers and decision strategies, our approach proposes a general framework for efficient hierarchical segmentation with statistical learning. We demonstrate that our method leads to a substantial improvement in segmentation accuracy.
Polyhydroxyalkanoate-based polymers—being ecofriendly, biosynthesizable, and economically viable and possessing a broad range of tunable properties—are currently being actively pursued as promising alternatives for petroleum-based plastics. The vast chemical complexity accessible within this class of polymers gives rise to challenges in the rational discovery of novel polymer chemistries for specific applications. Here, the burgeoning field of polymer informatics addresses this challenge via providing tools and strategies for accelerated property prediction and materials design via surrogate machine-learning models built on reliable past data. In this contribution, we use glass transition temperature T g as an example target property to demonstrate promise of the data-enabled route to accelerated learning of accurate structure–property mappings in PHA-based polymers. Our analysis uses a data set of experimentally measured T g values, polymer molecular weights, and a polydispersity index for PHA-based homo- and copolymers that was carefully assembled from the literature. A fingerprinting scheme that captures key properties based on topology, shape, and charge/polarity of specific chemical units or motifs forming the polymer backbone was devised to numerically represent the polymers. A validated statistical learning model is then developed to allow for a mapping of the polymer fingerprints onto the property space in a physically meaningful and reliable manner. Once developed, the model can not only rapidly predict the property of new PHA polymers but also provide uncertainties underlying the predictions. The model is further combined with an evolutionary-algorithm-based search strategy to efficiently identify multicomponent polymer compositions with a prespecified T g . While the present contribution is focused specifically on T g , the surrogate model development approach put forward here is general and can, in principle, be extended to a range of other properties.
A nuclear power plant is typically instrumented with a variety of sensors to continually monitor its variables, and their sensor’s measurements may be used to assess the plant state and initiate safety actions, if needed. Errors in sensor measurements, due to factors such as calibration drifts, critically affect such state assessments. Here, we address a problem of estimating sensor errors using physics-informed machine learning methods that use measurements collected under known plant conditions. For a given sensor, we propose an information fusion method that uses measurements from other sensors to estimate its output assuming it is error-free and provides its difference from an actual measurement as an error estimate. We present the ensemble of trees and support vector machine fusers, and evaluate their performance using measurements collected over an emulated test loop of a pressurized water reactor. The plant variables are related to each other through the underlying physical laws under inertial constraints that place bounds on their derivatives, which analytically justify the applicability of machine learning methods for computing these fusers. Under twenty scenarios, we assess their sensor error estimates for pressure sensors of the heat exchanger of a reactor’s primary coolant system. Multiple types of errors are captured by both fusers under externally induced calibration drifts, blockages, minor leaks and air gaps in sensing lines, and electromagnetic interference; the root mean square error of the estimation of error is under 2.2% percent of the maximum measurement. We present generalization equations, in the framework of statistical learning theory, for these methods that characterize the confidence probability that the estimation error is bounded by a specified parameter in future test scenarios.
RTRC and UMD have developed novel machine learning based methods under the ARPA-E DIFFERENTIATE program for rapid acceleration of hypothesis generation in complex architecture design spaces involving both discrete choices of component inclusion and interconnection and continuous parametric decisions. The project named Learning Enabled Network Synthesis (LENS) further demonstrated the developed methods on challenging electrical power converter design problems by identifying the most suitable circuit topologies and simultaneously selecting the most appropriate components to achieve optimized design of power converter with improved performances. We demonstrated that LENS could enable exploration of very large design space of circuit topologies and components by addressing the limitations of conventional design process in non-linear, high switching speed, multi-dimensional power converter design and optimization. The key innovation developed in LENS is the seamless integration of statistical learning and logical reasoning techniques and building on the individual strengths of these techniques for rapid hypothesis discovery. The main component of LENS comprises of: 1) Graph Reasoning Engine (GRE) to enforce composition rules that rapidly reject all discrete architectures that are composed incorrectly and generates an adaptive database of feasible designs which can be used by ML modules, 2) Graph Generative Learning module which is a deep neural network based generative model for graph architectures which can enable design space exploration beyond the dataset generated by the GRE, 3) Graph Reduced Order Model (ROM) for graph domains for accelerating computation of output metrics, and 4) Active learning and Rule Discovery module for sample efficient learning and extracting logical rules from the learned ML models which will be integrated in the GRE to enhance the filtering effectiveness. LENS approach can be applied to any design domains where designs can be represented as multi-attribute graphs. The LENS team integrated the various technical innovations listed above into an optimization pipeline and exercised the optimization pipeline on the converter design problem. The LENS project demonstrated that the developed AI/ML technologies can be used to generate novel converter circuits >45x faster than experts on chosen use-cases. This can enable faster design space exploration and identification of new designs which are not considered by experts due to the increasing design space complexity. This has significant potential impact on the public and energy needs of the country. It is currently estimated that 30% of all electrical powers generated passes through power converters. The future estimate is that 80% of all power generated would be passing through converters. LENS fills a critical gap in this space since by accelerating the design process the designers would be able to generate more efficient converters which can lead to significant energy savings for the country.
This report summarizes findings from research conducted for the “Exploration of the Poten tial for Artificial Intelligence and Machine Learning to Advance Low-Dose Radiation Biology Re search” (RadBio-AI) program, supported by the U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research, under Awards KP1601011/FWP CC121 and KP1601017/FWP CC121. The research reported here was undertaken in an effort to assess the potential of optimal measure transport methods as components within the larger scope of a com putational framework envisioned to support research in the radiation biology domain. Within this effort, our interest centered on enabling a unified generic framework where probabilistic modeling, inference, and statistical learning can be carried out for a wide range of data distributions. As described next in Section 1 (and in more detail in our original publication), optimal measure transport offers the possibility of such unified approach.
Characterizing the novel properties of layered van der Waals materials is key for their application in functional devices. A better understanding of this type of material requires correlative imaging of diverse nanoscale material properties. Within this class of materials, CuInP 2 S 6 (CIPS) has received a significant degree of interest due to its ionically mediated room temperature ferroelectricity. Moreover, it is possible to form stable self-assembled heterostructures of ferroelectric CuInP 2 S 6 (CIPS) and non-ferroelectric (i.e., lacking Cu) In 4/3 P 2 S 6 (IPS) phases, by controlling the targeted composition and kinetics of synthesis. In this work, we present a correlative nanometric imaging study of the phononic modes and piezoelectricity of the phase-separated thin heteroepitaxial CIPS/IPS flakes. Here, we show that it is possible to isolate the different phononic modes of the two phases by spatially correlating them with their distinct ferroelectric behavior. The coupling of our experimental data with unsupervised learning statistical methods enables unraveling specific Raman peaks that are characteristic of each chemical phase (CIPS and IPS) present in the composite sample, discarding the less significant ones.
Abstract Background The findings of the 2023 AAPM Grand Challenge on Deep Generative Modeling for Learning Medical Image Statistics are reported in this Special Report. Purpose The goal of this challenge was to promote the development of deep generative models for medical imaging and to emphasize the need for their domain‐relevant assessments via the analysis of relevant image statistics. Methods As part of this Grand Challenge, a common training dataset and an evaluation procedure was developed for benchmarking deep generative models for medical image synthesis. To create the training dataset, an established 3D virtual breast phantom was adapted. The resulting dataset comprised about 108 000 images of size 512 512. For the evaluation of submissions to the Challenge, an ensemble of 10 000 DGM‐generated images from each submission was employed. The evaluation procedure consisted of two stages. In the first stage, a preliminary check for memorization and image quality (via the Fréchet Inception Distance [FID]) was performed. Submissions that passed the first stage were then evaluated for the reproducibility of image statistics corresponding to several feature families including texture, morphology, image moments, fractal statistics, and skeleton statistics. A summary measure in this feature space was employed to rank the submissions. Additional analyses of submissions was performed to assess DGM performance specific to individual feature families, the four classes in the training data, and also to identify various artifacts. Results Fifty‐eight submissions from 12 unique users were received for this Challenge. Out of these 12 submissions, 9 submissions passed the first stage of evaluation and were eligible for ranking. The top‐ranked submission employed a conditional latent diffusion model, whereas the joint runners‐up employed a generative adversarial network, followed by another network for image superresolution. In general, we observed that the overall ranking of the top 9 submissions according to our evaluation method (i) did not match the FID‐based ranking, and (ii) differed with respect to individual feature families. Another important finding from our additional analyses was that different DGMs demonstrated similar kinds of artifacts. Conclusions This Grand Challenge highlighted the need for domain‐specific evaluation to further DGM design as well as deployment. It also demonstrated that the specification of a DGM may differ depending on its intended use.
DEEPEN stands for DE-risking Exploration of geothermal Plays in magmatic ENvironments. As part of the development of the DEEPEN 3D play fairway analysis (PFA) methodology for magmatic plays (conventional hydrothermal, superhot EGS, and supercritical), weights needed to be developed for use in the weighted sum of the different favorability index models produced from geoscientific exploration datasets. This GDR submission includes those weights. The weighting was done using two different approaches: one based on expert opinions, and one based on statistical learning. The weights are intended to describe how useful a particular exploration method is for imaging each component of each play type. They may be adjusted based on the characteristics of the resource under investigation, knowledge of the quality of the dataset, or simply to reduce the impact a single dataset has on the resulting outputs. Within the DEEPEN PFA, separate sets of weights are produced for each component of each play type, since exploration methods hold different levels of importance for detecting each play component, within each play type. The weights for conventional hydrothermal systems were based on the average of the normalized weights used in the DOE-funded PFA projects that were focused on magmatic plays. This decision was made because conventional hydrothermal plays are already well-studied and understood, and therefore it is logical to use existing weights where possible. In contrast, a true PFA has never been applied to superhot EGS or supercritical plays, meaning that exploration methods have never been weighted in terms of their utility in imaging the components of these plays. To produce weights for superhot EGS and supercritical plays, two different approaches were used: one based on expert opinion and the analytical hierarchy process (AHP), and another using a statistical approach based on principal component analysis (PCA). The weights are intended to provide standardized sets of weights for each play type in all magmatic geothermal systems. Two different approaches were used to investigate whether a more data-centric approach might allow new insights into the datasets, and also to analyze how different weighting approaches impact the outcomes. The expert/AHP approach involved using an online tool (https://bpmsg.com/ahp/) with built-in forms to make pairwise comparisons which are used to rank exploration methods against one-another. The inputs are then combined in a quantitative way, ultimately producing a set of consensus-based weights. To minimize the burden on each individual participant, the forms were completed in group discussions. While the group setting means that there is potential for some opinions to outweigh others, it also provides a venue for conversation to take place, in theory leading the group to a more robust consensus then what can be achieved on an individual basis. This exercise was done with two separate groups: one consisting of U.S.-based experts, and one consisting of Iceland-based experts in magmatic geothermal systems. The two sets of weights were then averaged to produce what we will from here on refer to as the "expert opinion-based weights," or "expert weights" for short. While expert opinions allow us to include more nuanced information in the weights, expert opinions are subject to human bias. Data-centric or statistical approaches help to overcome these potential human biases by focusing on and drawing conclusions from the data alone. More information on this approach along with the dataset used to produce the statistical weights may be found in the linked dataset below.
A workshop on artificial intelligence and machine learning (AI/ML) for advanced reactors (AR) was held October 5-6, 2021. The workshop was to be attended in-person at ANL but COVID restrictions forced the workshop to go virtual. The objectives of the workshop were to identify the most promising AI/ML opportunities for improving advanced reactor design, optimizing plant performance, and enhancing economic competitiveness and to develop an understanding of the scientific, engineering and licensing challenges facing their application. The workshop planning committee included GAIN, EPRI and NEI and members of three national laboratories (ANL, INL, and ORNL). The workshop was attended by more than 200 individuals representing academic and scientific institutions and the nuclear power industry. The definition put forth for an AI/ML system was one that perceives its environment and takes actions that maximize its chance of achieving its goals. In this report AI/ML refers to next generation algorithms that include deep learning, statistical analysis and data analytics and associated scientific computing and their potential application to the design, licensing, operation and maintenance of ARs. These methods typically incorporate models built from process data and may also include data generated by simulations that represent the behavior of a system. The workshop was organized in response to the growing interest in application of AI/ML for improving the economic competitiveness of nuclear energy. Increasingly more resources are being allocated to investigating the benefits of AI/ML methods. The DOE created the Artificial Intelligence & Technology Office to promote their development. And within the Office of Nuclear Energy, resources have been allocated to explore and understand the potential benefits of AI/ML. Additionally, the national laboratories are strategically positioned with DOE computing facilities such as Summit, Perlmutter, Aurora and Frontier that support large-scale simulations, hybrid HPC models with AI surrogates, and the exploration of new types of generative models emerging from multi-model data streams and sources. The workshop was organized with members of the AR community to understand the effort and to identify the level of interest and progress in this emerging technology. The workshop discussions focused on identifying opportunities for AI/ML across diverse areas of the nuclear industry and identifying current scientific and engineering challenges for advanced reactors that might be addressed through transformational uses of AI/ML. Discussion panels focused on four high-interest technical domains for advanced reactors: design, maintenance and operations, energy storage, and materials. The results of those discussions are summarized in this report. This includes opportunities that were identified for exploiting AI techniques and methods to improve the efficacy and efficiency of reactor analysis and to improve the operation and optimization of advanced reactors. Advanced reactor developers expressed an interest in learning more about AI/ML methods and their application. This included understanding whether ML methods can provide an advantage over existing nonlinear data regression methods for collapsing high-fidelity simulation results into faster running models. A consensus emerged that AR advances planned for the next decade will benefit from the use of AI/ML tools. The need exists to understand and model complex systems across length scales and modalities. AI/ML is a tool for discovery that can yield a set of engineering principles for use by nuclear engineers, licensing bodies, and operators to solve problems in plant design, safety analyses, autonomous operation, and predictive maintenance. While AI/ML represents a new set of tools, an awareness by the nuclear community of the full potential is still in the early stages so there is a need to increase awareness. It appears that the wide-spread adoption of AI/ML tools for ARs would be facilitated by future educational workshops that describe foundational methods and capabilities and describe successful applications.