Search NASA⌕ Search

SEARCH · Search NASA

Results for “spatial optimization”

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.

At least 631 records · Page 35

MO-F-CAMPUS-I-04: Characterization of Fan Beam Coded Aperture Coherent Scatter Spectral Imaging Methods for Differentiation of Normal and Neoplastic Breast Structures

Purpose: This study intends to characterize the spectral and spatial resolution limits of various fan beam geometries for differentiation of normal and neoplastic breast structures via coded aperture coherent scatter spectral imaging techniques. In previous studies, pencil beam raster scanning methods using coherent scatter computed tomography and selected volume tomography have yielded excellent results for tumor discrimination. However, these methods don’t readily conform to clinical constraints; primarily prolonged scan times and excessive dose to the patient. Here, we refine a fan beam coded aperture coherent scatter imaging system to characterize the tradeoffs between dose, scan time and image quality for breast tumor discrimination. Methods: An X-ray tube (125kVp, 400mAs) illuminated the sample with collimated fan beams of varying widths (3mm to 25mm). Scatter data was collected via two linear-array energy-sensitive detectors oriented parallel and perpendicular to the beam plane. An iterative reconstruction algorithm yields images of the sample’s spatial distribution and respective spectral data for each location. To model in-vivo tumor analysis, surgically resected breast tumor samples were used in conjunction with lard, which has a form factor comparable to adipose (fat). Results: Quantitative analysis with current setup geometry indicated optimal performance for beams up to 10mm wide, with wider beams producing poorer spatial resolution. Scan time for a fixed volume was reduced by a factor of 6 when scanned with a 10mm fan beam compared to a 1.5mm pencil beam. Conclusion: The study demonstrates the utility of fan beam coherent scatter spectral imaging for differentiation of normal and neoplastic breast tissues has successfully reduced dose and scan times whilst sufficiently preserving spectral and spatial resolution. Future work to alter the coded aperture and detector geometries could potentially allow the use of even wider fans, thereby making coded aperture coherent scatter imaging a clinically viable method for breast cancer detection. United States Department of Homeland Security; Duke University Medical Center - Department of Radiology; Carl E Ravin Advanced Imaging Laboratories; Duke University Medical Physics Graduate Program.

60 APPLIED LIFE SCIENCES↗

SU-D-BRC-01: An Automatic Beam Model Commissioning Method for Monte Carlo Simulations in Pencil-Beam Scanning Proton Therapy

Purpose: Monte Carlo (MC) simulation is typically regarded as the most accurate dose calculation method for proton therapy. Yet for real clinical cases, the overall accuracy also depends on that of the MC beam model. Commissioning a beam model to faithfully represent a real beam requires finely tuning a set of model parameters, which could be tedious given the large number of pencil beams to commmission. This abstract reports an automatic beam-model commissioning method for pencil-beam scanning proton therapy via an optimization approach. Methods: We modeled a real pencil beam with energy and spatial spread following Gaussian distributions. Mean energy, and energy and spatial spread are model parameters. To commission against a real beam, we first performed MC simulations to calculate dose distributions of a set of ideal (monoenergetic, zero-size) pencil beams. Dose distribution for a real pencil beam is hence linear superposition of doses for those ideal pencil beams with weights in the Gaussian form. We formulated the commissioning task as an optimization problem, such that the calculated central axis depth dose and lateral profiles at several depths match corresponding measurements. An iterative algorithm combining conjugate gradient method and parameter fitting was employed to solve the optimization problem. We validated our method in simulation studies. Results: We calculated dose distributions for three real pencil beams with nominal energies 83, 147 and 199 MeV using realistic beam parameters. These data were regarded as measurements and used for commission. After commissioning, average difference in energy and beam spread between determined values and ground truth were 4.6% and 0.2%. With the commissioned model, we recomputed dose. Mean dose differences from measurements were 0.64%, 0.20% and 0.25%. Conclusion: The developed automatic MC beam-model commissioning method for pencil-beam scanning proton therapy can determine beam model parameters with satisfactory accuracy.

60 APPLIED LIFE SCIENCES↗

Enhanced deep neural networks with transfer learning for distribution LMP considering load and PV uncertainties

As the flexibility of generation and demand increases in distribution systems, the residential loads are emerging as a promising means to participate in demand response and the transactive energy market. Market pricing is an instrumental mechanism for the distribution system operator to exploit the full potential of the flexible resources. The distribution locational marginal price (DLMP) can be used to guide the residential load consumption. This type of market signal helps the distribution system operator to optimize the scheduling of all resources while satisfying related network constraints through a day-ahead market. However, solving the optimization problem for large-scale systems can be computationally expensive. To address the scalability and practicability limitations of the DLMP framework, a learning-based approach is proposed in this paper to complement the day-ahead distribution market framework. Here, the proposed approach combines long short-term memory and transfer learning to develop deep neural network that can capture the spatial–temporal correlation of the input data. The model can determine the optimal DLMP for each node in a distribution system without the system parameters required to formulate the optimization problem. Testing results on IEEE 33-bus and 123-bus systems show that the proposed approach can generate a comparable DLMP against the optimization solutions.

24 POWER TRANSMISSION AND DISTRIBUTION↗

Distributed Optimization for Nonrigid Nano-Tomography

Resolution level and reconstruction quality in nano-computed tomography (nano-CT) are in part limited by the stability of microscopes, because the magnitude of mechanical vibrations during scanning becomes comparable to the imaging resolution, and the ability of the samples to resist radiation induced deformations during data acquisition. In such cases, there is no incentive in recovering the sample state at different time steps like in time-resolved reconstruction methods, but instead the goal is to retrieve a single reconstruction at the highest possible spatial resolution and without any imaging artifacts. Here we propose a distributed optimization solver for tomographic imaging of samples at the nanoscale. Our approach solves the tomography problem jointly with projection data alignment, nonrigid sample deformation correction, and regularization. Projection data consistency is regulated by dense optical flow estimated by Farneback's algorithm, leading to sharp sample reconstructions with less artifacts. Synthetic data tests show robustness of the method to Poisson and low-frequency background noise. We accelerated the solver on multi-GPU systems and validated the method on three nano-imaging experimental data sets.

97 MATHEMATICS AND COMPUTING↗

Journey Over Destination: Differentiable Sensor Placement Enhances Generalization [Poster]

The challenge of reconstructing spatial fields that change over time from limited sensor data has been a focal point for many research studies. Various machine learning methods have been used in attempts to address this complex issue, including convolutional neural networks. All of the proposed methods share a common requirement that the user needs to manually determine the sensor positions. This requirement remains a limiting factor in the ongoing quest for efficient learning and accurate field reconstruction. This study aims to present a method that enables a model to optimize sensor positions via backpropagation, thereby facilitating the model’s exploration of the spatial domain and enhancing sensor positioning effectively. Indexing naturally incorporates discrete decisions. This operation is nondifferentiable which is a requirement for the application of gradient-based optimization methods. We showcased its effectiveness by training an attention-based neural network, which achieved top-tier performance on two separate datasets. To our knowledge, this represents the first fully end-to-end differentiable workflow for enhancing sensor placement within a neural network model.

58 GEOSCIENCES↗

Bilevel Parameter Optimization for Nonlocal Image Denoising Models

We propose a bilevel optimization approach for the determination of parameters in nonlocal image denoising. We consider both spatial weights in front of the fidelity term, as well as weights within the kernel of the nonlocal operator. In both cases we investigate the differentiability of the solution operator in function spaces and derive a first order optimality system that characterizes local minima. For the numerical solution of the problems, we propose a second-order optimization algorithm in combination with a finite element discretization of the nonlocal denoising models and a computational strategy for the solution of the resulting dense linear systems. Several experiments are run in order to show the suitability of our approach.

97 MATHEMATICS AND COMPUTING↗

Constrained or unconstrained? Neural-network-based equation discovery from data

Throughout many fields, practitioners often rely on differential equations to model systems. Yet, for many applications, the theoretical derivation of such equations and/or the accurate resolution of their solutions may be intractable. Instead, recently developed methods, including those based on parameter estimation, operator subset selection, and neural networks, allow for the data-driven discovery of both ordinary and partial differential equations (PDEs), on a spectrum of interpretability. The success of these strategies is often contingent upon the correct identification of representative equations from noisy observations of state variables and, as importantly and intertwined with that, the mathematical strategies utilized to enforce those equations. Specifically, the latter has been commonly addressed via unconstrained optimization strategies. Representing the PDE as a neural network, we propose to discover the PDE (or the associated operator) by solving a constrained optimization problem and using an intermediate state representation similar to a physics-informed neural network (PINN). The objective function of this constrained optimization problem promotes matching the data, while the constraints require that the discovered PDE is satisfied at a number of spatial collocation points. We present a penalty method and a widely used trust-region barrier method to solve this constrained optimization problem, and we compare these methods on numerical examples. Our results on several example problems demonstrate that the latter constrained method outperforms the penalty method, particularly for higher noise levels or fewer collocation points. This work motivates further exploration into using sophisticated constrained optimization methods in scientific machine learning, as opposed to their commonly used, penalty-method or unconstrained counterparts. For both of these methods, we solve these discovered neural network PDEs with classical methods, such as finite difference methods, as opposed to PINNs-type methods relying on automatic differentiation. Here, we briefly highlight how simultaneously fitting the data while discovering the PDE improves the robustness to noise and other small, yet crucial, implementation details.

Data-driven discovery↗

Real time time-dependent density functional theory using higher order finite-element methods

We present a computationally efficient approach to solve the time-dependent Kohn-Sham equations in real time using higher order finite-element spatial discretization, applicable to both pseudopotential and all-electron calculations. To this end, we develop an a priori mesh adaption technique, based on the semidiscrete (discrete in space but continuous in time) error estimate on the time-dependent Kohn-Sham orbitals, to construct an efficient finite-element discretization. Subsequently, we obtain the full-discrete error estimate to guide our choice of the time step. We employ spectral finite elements along with special reduced order quadrature to render the overlap matrix diagonal, thereby simplifying the inversion of the overlap matrix that features in the evaluation of the discrete time-evolution operator. We use the second-order Magnus operator as the time-evolution operator, wherein the action of the discrete Magnus operator, expressed as exponential of a matrix, on the Kohn-Sham orbitals is obtained efficiently through an adaptive Lanczos iteration. We observe close to optimal rates of convergence of the dipole moment with respect to spatial and temporal discretization, for both pseudopotential and all-electron calculations. We demonstrate a staggering 100-fold reduction in the computational time afforded by higher order finite elements over linear finite elements, for both pseudopotential and all-electron calculations. Further, for similar level of accuracy, we obtain significant computational savings by our approach as compared to state-of-the-art finite-difference methods. Here, we also demonstrate the competence of higher order finite elements for all-electron benchmark systems. Lastly, we observe good parallel scalability of the proposed method on many hundreds of processors.

36 MATERIALS SCIENCE↗

rlmolecule: A library for general-purpose material and molecular optimization using AlphaZero-style reinforcement learning

This library includes in-progress code for the optimization of materials and molecules using reinforcement learning. This project seeks to develop a generalized machine learning approach for optimizing targeted, and often complex, functional properties over the space of possible material structures, allowing faster identification of promising candidates. Electronic and transport properties of materials are determined by their molecular structure (for molecules) or their crystal structure (for in-organic crystalline materials). The atomic composition, bonding, and 3-D-spatial arrangement in these systems therefore represent a vast, discontinuous search space for optimal candidates. This library implements AlphaZero style reinforcement learning applied to two worked examples: the optimization of stable organic radicals for redox flow batteries, and the determination of stable electrolyte species for solid state batteries. The current code contains proof-of-concept results run on the Eagle HPC system, while the completed package will be a flexible tool for material optimization across of number of energy-relevant fields, which leverages cloud compute resources for the majority of the computational heavy lifting.

Biagioni, David↗

Symmetric Phase-Only Filtering in Particle-Image Velocimetry

Symmetrical phase-only filtering (SPOF) can be exploited to obtain substantial improvements in the results of data processing in particle-image velocimetry (PIV). In comparison with traditional PIV data processing, SPOF PIV data processing yields narrower and larger amplitude correlation peaks, thereby providing more-accurate velocity estimates. The higher signal-to-noise ratios associated with the higher amplitude correlation peaks afford greater robustness and reliability of processing. SPOF also affords superior performance in the presence of surface flare light and/or background light. SPOF algorithms can readily be incorporated into pre-existing algorithms used to process digitized image data in PIV, without significantly increasing processing times. A summary of PIV and traditional PIV data processing is prerequisite to a meaningful description of SPOF PIV processing. In PIV, a pulsed laser is used to illuminate a substantially planar region of a flowing fluid in which particles are entrained. An electronic camera records digital images of the particles at two instants of time. The components of velocity of the fluid in the illuminated plane can be obtained by determining the displacements of particles between the two illumination pulses. The objective in PIV data processing is to compute the particle displacements from the digital image data. In traditional PIV data processing, to which the present innovation applies, the two images are divided into a grid of subregions and the displacements determined from cross-correlations between the corresponding sub-regions in the first and second images. The cross-correlation process begins with the calculation of the Fourier transforms (or fast Fourier transforms) of the subregion portions of the images. The Fourier transforms from the corresponding subregions are multiplied, and this product is inverse Fourier transformed, yielding the cross-correlation intensity distribution. The average displacement of the particles across a subregion results in a displacement of the correlation peak from the center of the correlation plane. The velocity is then computed from the displacement of the correlation peak and the time between the recording of the two images. The process as described thus far is performed for all the subregions. The resulting set of velocities in grid cells amounts to a velocity vector map of the flow field recorded on the image plane. In traditional PIV processing, surface flare light and bright background light give rise to a large, broad correlation peak, at the center of the correlation plane, that can overwhelm the true particle- displacement correlation peak. This has made it necessary to resort to tedious image-masking and background-subtraction procedures to recover the relatively small amplitude particle-displacement correlation peak. SPOF is a variant of phase-only filtering (POF), which, in turn, is a variant of matched spatial filtering (MSF). In MSF, one projects a first image (denoted the input image) onto a second image (denoted the filter) as part of a computation to determine how much and what part of the filter is present in the input image. MSF is equivalent to cross-correlation. In POF, the frequency-domain content of the MSF filter is modified to produce a unitamplitude (phase-only) object. POF is implemented by normalizing the Fourier transform of the filter by its magnitude. The advantage of POFs is that they yield correlation peaks that are sharper and have higher signal-to-noise ratios than those obtained through traditional MSF. In the SPOF, these benefits of POF can be extended to PIV data processing. The SPOF yields even better performance than the POF approach, which is uniquely applicable to PIV type image data. In SPOF as now applied to PIV data processing, a subregion of the first image is treated as the input image and the corresponding subregion of the second image is treated as the filter. The Fourier transforms from both the firs and second- image subregions are normalized by the square roots of their respective magnitudes. This scheme yields optimal performance because the amounts of normalization applied to the spatial-frequency contents of the input and filter scenes are just enough to enhance their high-spatial-frequency contents while reducing their spurious low-spatial-frequency content. As a result, in SPOF PIV processing, particle-displacement correlation peaks can readily be detected above spurious background peaks, without need for masking or background subtraction.

Wemet, Mark P.↗

Development of a novel high quantum efficiency MV x-ray detector for image-guided radiotherapy: A feasibility study

Purpose: To develop a new scintillating fiber-based electronic portal imaging device (EPID) with a high quantum efficiency (QE) while preserving an adequate spatial resolution. Methods: Two prototypes were built: one with a single pixel readout and the other with an active matrix flat-panel imager (AMFPI) for readout. The energy conversion layer of both prototypes was made of scintillating fiber layers interleaved with corrugated lead sheets to form a honeycomb pattern. The scintillating fibers have a diameter of 1 mm and the distance between the centers of neighboring fibers on the same layer is 1.35 mm. The layers have 1.22 mm spacing between them. The energy conversion layer has a thickness of 2 cm. The modulation transfer function (MTF), antiscatter properties and sensitivity of the detector with a single pixel readout were measured using a 6-MV beam on a LINAC machine. In addition, a Monte Carlo simulation was conducted to calculate the zero-frequency detective quantum efficiency (DQE(0)) of the proposed detector with an active matrix flat-panel imager for readout. Results: The DQE(0) of the proposed detector can be 11.5%, which is about an order of magnitude higher than that of current EPIDs. The frequency of 50% modulation (f 50 ) of the measured MTF is 0.2 mm –1 at 6 MV, which is comparable to that of video-based EPIDs. The scatter to primary ratio (SPR) measured with the detector at 10 cm air gap and 20 x 20 cm 2 field size is approximately 30% lower than that of ionization chamber–based detectors with a comparable QE. The detector noise which includes the x-ray quantum noise and absorption noise is much larger than the electronic noise per pixel of the flat-panel imager at a dose of less than two LINAC pulses. Thus, the proposed detector is quantum noise limited down to very low doses (~a couple of radiation pulses of the LINAC). A proof-of-concept image has been obtained using a 6-MV beam. Conclusions: This work indicates that by using scintillating fibers and lead layers it is possible to increase the thickness of the detecting materials, and therefore the QE or the DQE(0) of the detector, while maintaining an adequate spatial resolution for MV x-ray imaging. Due to the use of lead as the spacing material, the new detector also has antiscatter property, which will help improve the signal-to-noise ratio of the images. Further investigation to optimize the design of the detector and achieve a better combination of DQE and spatial resolution is warranted.

62 RADIOLOGY AND NUCLEAR MEDICINE↗

Spatio-Temporal Patterns of Optimal Landsat Data for Burn Severity Index Calculations: Implications for High Northern Latitudes Wildfire Research

Satellite remote sensing has been widely used for the evaluation of wildfire burn severity in various ecosystems. While a variety of remote sensing-based burn severity indices have been developed, the Landsat-based differ-enced Normalized Burn Ratio (dNBR) presents the most widely-used approach to burn severity assessment for fire research and management. Although dNBR-based approaches have been continuously updated, including the development of the relative dNBR (RdNBR) and the Relativized Burn Ratio (RBR), one key obstacle for the reliable applications of the burn severity indices in the high northern latitudes has not been adequately addressed. Specifically, optimal Landsat image pairs are very rarely available for the calculation of the burn severity indices in certain regions of the high northern latitudes (HNL), resulting in the burn severity index outputs calculated for a large number of wildfires being considerably affected by non-fire-related factors. The suboptimal selection of image pairs may have partially contributed to the lack of consistency in the performance of the burn severity indices in the HNL regions. In this paper, we systematically evaluated the impacts of sub-optimal image pairs on signal stability through two sets of analyses conducted at different spatial scales. First, at the regional scale, we examined the burn severity indices calculated for a selection of wildfires in Alaska. Here, we demonstrate the inconsistent performance of the same index calculated based on Landsat data that are typically considered as “optimal”. In addition, we show that the calculated indices may be more prone to the negative impacts of these inconsistencies, thus leading to reduced reliability, when there is limited Landsat data availability. Second, at the continental scale, we show that certain areas in the HNL, especially Alaska during the pre-2000 era, are subject to the potentially strong negative impact associated with the limited data availability. Through systematically analyzing this issue, we hope to not only divert more attention to it but also to provide potential solutions, based on which further improvements that may be of particular importance for research on Arctic wildfires could be made.

Dong Chen↗

Loading and Regeneration Analysis of a Diesel Particulate Filter with a Radio Frequency-Based Sensor

Accurate knowledge of diesel particulate filter (DPF) loading is critical for robust and efficient operation of the combined engine-exhaust aftertreatment system. Furthermore, upcoming on-board diagnostics regulations require on-board technologies to evaluate the status of the DPF. This work describes the application of radio frequency (RF) based sensing techniques to accurately measure DPF soot levels and the spatial distribution of the accumulated material. A 1.9L GM turbo diesel engine and a DPF with an RF-sensor were studied. Direct comparisons between the RF measurement and conventional pressure-based methods were made. Further analysis of the particulate matter loading rates was obtained with a mass-based soot emission measurement instrument (TEOM). Comparison with pressure drop measurements show the RF technique is unaffected by exhaust flow variations and exhibits a high degree of sensitivity to DPF soot loading and good dynamic response. Additional computational and experimental work further illustrates the spatial resolution of the RF measurements. Based on the experimental results, the RF technique shows significant promise for improving DPF control enabling optimization of the combined engine-aftertreatment system for improved fuel economy and extended DPF service life.

33 ADVANCED PROPULSION SYSTEMS↗

OPTIMIZATION-BASED CONSTITUTIVE PARAMETER IDENTIFICATION FROM SPARSE TAYLOR CYLINDER DATA

The classic Taylor impact test imparts temporally and spatially varying fields of strain, strain rate, and temperature through the specimen. It is possible to exploit this complexity to directly identify constitutive model parameters from the deformed shape of the specimen. Where prior investigators have employed various mathematical fitting methods to identify or improve strength model parameters from Taylor cylinder profiles, we extend the method to employ a multi-objective genetic optimization algorithm to minimize the cylinder profile errors simultaneously on three cylinders impacted at different velocities. No experimental data other than the three Taylor cylinders is employed in developing the constitutive model parameter set, and generic starting coefficients are employed. To validate the accuracy of the resulting coefficients, both split Hopkinson pressure bar and axisymmetric expanding ring tests were conducted and compared to the resultant Johnson-Cook strength model. The derived strength model agreed well with experimental data available to date. Further work is necessary to evaluate the range of rates and temperatures over which parameters derived by this method may be applied.

97 MATHEMATICS AND COMPUTING↗

Intelligent automated surface grid generation

The goal of our research is to produce a flexible, general grid generator for automated use by other programs, such as numerical optimizers. The current trend in the gridding field is toward interactive gridding. Interactive gridding more readily taps into the spatial reasoning abilities of the human user through the use of a graphical interface with a mouse. However, a sometimes fruitful approach to generating new designs is to apply an optimizer with shape modification operators to improve an initial design. In order for this approach to be useful, the optimizer must be able to automatically grid and evaluate the candidate designs. This paper describes and intelligent gridder that is capable of analyzing the topology of the spatial domain and predicting approximate physical behaviors based on the geometry of the spatial domain to automatically generate grids for computational fluid dynamics simulators. Typically gridding programs are given a partitioning of the spatial domain to assist the gridder. Our gridder is capable of performing this partitioning. This enables the gridder to automatically grid spatial domains of wide range of configurations.

Yao, Ke-Thia↗

The ALS OSMS: Optical Surface Measuring System for high accuracy two-dimensional slope metrology with state-of-the-art x-ray mirrors

We report that to preserve the brightness and coherence of x-rays produced by diffraction-limited-storage-ring (DLSR) and free-electron laser (FEL) light sources, beamline optics must have unprecedented quality. For example, in the case of the most advanced beamlines for the DLSR source under development at the Advanced Light Source (ALS), the ALS-U, we need highly curved x-ray mirrors with surface slope tolerances better than 50-100 nrad (root-mean-square, rms). At the ALS X-Ray Optics Lab (XROL), we are working on the development of a new Optical Surface Measuring System (OSMS) with the required measurement accuracy. The OSMS is capable for the two-dimensional (2D) surface slope metrology over the spatial range from the sub-mm scale to the clear aperture. Usage of different arrangements of the OSMS sensors allows measuring the mirrors in the face-up or side-facing orientation, corresponding to the beamline application. The OSMS translation system and data acquisition software are designed to support multi scan measurement runs optimized for automatic suppression and compensation of instrumental drifts and major angular and spatial systematic errors. Here, we discuss the recent results of the OSMS research and development project. We provide details of the OSMS design and describe results of experimental performance tests of the gantry system. In particular, we show that the system is capable for measurement repeatability with strongly curved mirrors on the level of 20 nrad (rms). The high angular resolution of the OSMS rotational tip-tilt stage is adequate for implementation of instrumental calibration with using the mirror under test as a reference. The achieved measuring accuracy is demonstrated via comparison to metrology with the carefully calibrated Developmental Long Trace Profiler, also available at the XROL.

47 OTHER INSTRUMENTATION↗

An optimized finite-difference scheme for wave propagation problems

Two fully-discrete finite-difference schemes for wave propagation problems are presented, a maximum-order scheme and an optimized (or spectral-like) scheme. Both combine a seven-point spatial operator and an explicit six-stage time-march method. The maximum-order operator is fifth-order in space and is sixth-order in time for a linear problem with periodic boundary conditions. The phase and amplitude errors of the schemes obtained using Fourier analysis are given and compared with a second-order and a fourth-order method. Numerical experiments are presented which demonstrate the usefulness of the schemes for a range of problems. For some problems, the optimized scheme leads to a reduction in global error compared to the maximum-order scheme with no additional computational expense.

Zingg, D. W.↗

Sparse coding of pathology slides compared to transfer learning with deep neural networks

Background Histopathology images of tumor biopsies present unique challenges for applying machine learning to the diagnosis and treatment of cancer. The pathology slides are high resolution, often exceeding 1GB, have non-uniform dimensions, and often contain multiple tissue slices of varying sizes surrounded by large empty regions. The locations of abnormal or cancerous cells, which may constitute a small portion of any given tissue sample, are not annotated. Cancer image datasets are also extremely imbalanced, with most slides being associated with relatively common cancers. Since deep representations trained on natural photographs are unlikely to be optimal for classifying pathology slide images, which have different spectral ranges and spatial structure, we here describe an approach for learning features and inferring representations of cancer pathology slides based on sparse coding. Results We show that conventional transfer learning using a state-of-the-art deep learning architecture pre-trained on ImageNet (RESNET) and fine tuned for a binary tumor/no-tumor classification task achieved between 85% and 86% accuracy. However, when all layers up to the last convolutional layer in RESNET are replaced with a single feature map inferred via a sparse coding using a dictionary optimized for sparse reconstruction of unlabeled pathology slides, classification performance improves to over 93%, corresponding to a 54% error reduction. Conclusions We conclude that a feature dictionary optimized for biomedical imagery may in general support better classification performance than does conventional transfer learning using a dictionary pre-trained on natural images.

59 BASIC BIOLOGICAL SCIENCES↗