Publications

Preprints

Journal Papers

1. K. Falahkheirkhah, K. Yeh, S. Mittal, L. Pfister, and R. Bhargava, “Deep Learning-Based Protocols To Enhance Infrared Imaging Systems,” Chemometrics and Intelligent Laboratory Systems, vol. 217, no. nil, 2021.
   • Abstract
   Infrared (IR) spectroscopic imaging provides both morphologic and chemical detail; however, obtaining this extensive spectral-spatial information requires the ability to rapidly record high-quality data. Discrete frequency infrared (DFIR) imaging using a point scanning microscope strikes a balance between data quality and acquisition speed that, in principle, can further be aided by computational methods. Here, we report a deep learning-based framework to complement the process of data acquisition and information extraction. First, we introduce a convolutional neural network (CNN) to leverage both spatial and spectral information for segmenting data into informative sub-classes, which we call the IR-SEG network. We show that this framework increases accuracy by using approximately half of the features used in the typical pixel-wise classification of IR data. Second, we present a generative adversarial network (GAN)-based approach to reconstruct full data sets with low loss in the information from incomplete spatial and spectral data recording. Termed IR-REC, this approach is shown to speed up data acquisition by up to 20-fold for typical biomedical samples. In addition to enhancing the speed and quality of data, we also propose a method to utilize complementary morphologic detail to estimate the spatial details of a single band IR image beyond the diffraction limit. Finally, we discuss potential pitfalls and new opportunities that can be addressed by developing these methods further. Together, these deep learning techniques provide new capabilities for IR imaging to extract better quality information faster.
   • Link to Publisher
   • BibTeX

   @article{Falahkheirkhah2021-Deep-Learn,
     author = {Falahkheirkhah, Kianoush and Yeh, Kevin and Mittal, Shachi and Pfister, Luke and Bhargava, Rohit},
     title = {Deep Learning-Based Protocols To Enhance Infrared Imaging
                       Systems},
     journal = {Chemometrics and Intelligent Laboratory Systems},
     volume = {217},
     number = {nil},
     pages = {104390},
     year = {2021},
     doi = {10.1016/j.chemolab.2021.104390},
     url = {https://doi.org/10.1016/j.chemolab.2021.104390}
   }
   

   
   
2. L. Pfister, R. Bhargava, Y. Bresler, and S. Carney, “Composition-Aware Spectroscopic Tomography,” Inverse Problems, 2020.
   • Download
   • Link to Publisher
   • BibTeX

   @article{Pfister2020-Compos-Aware,
     author = {Pfister, Luke and Bhargava, Rohit and Bresler, Yoram and Carney, Scott},
     title = {Composition-Aware Spectroscopic Tomography},
     journal = {Inverse Problems},
     year = {2020},
     doi = {10.1088/1361-6420/abb767},
     url = {https://doi.org/10.1088/1361-6420/abb767}
   }
   

   
   
3. B. Wen, S. Ravishankar, L. Pfister, and Y. Bresler, “Transform Learning for Magnetic Resonance Image Reconstruction: From Model-Based Learning To Building Neural Networks,” IEEE Signal Processing Magazine, vol. 37, no. 1, 2020.
   • Abstract
   Magnetic resonance imaging (MRI) is widely used in clinical practice for visualizing both biological structure and function, but its use has been traditionally limited by its slow data acquisition. Recent advances in compressed sensing (CS) techniques for MRI that exploit sparsity models of images reduce acquisition time while maintaining high image quality. Whereas classical CS assumes the images are sparse in a known analytical dictionary or transform domain, methods that use learned image models for reconstruction have become popular in recent years. The model could be learned from a dataset and used for reconstruction or learned simultaneously with the reconstruction, a technique called blind CS (BCS). While the well-known synthesis dictionary model has been exploited for MRI reconstruction, recent advances in transform learning (TL) provide an efficient alternative framework for sparse modeling in MRI. TL-based methods enjoy numerous advantages including exact sparse coding, transform update, and clustering solutions, cheap computation, and convergence guarantees, and provide high quality results in MRI as well as in other inverse problems compared to popular competing methods. This paper provides a review of key works in MRI reconstruction from limited data, with focus on the recent class of TL-based reconstruction methods. A unified framework for incorporating various TL-based models is presented. We discuss the connections between transform learning and convolutional or filterbank models and corresponding multi-layer extensions, as well as connections to unsupervised and supervised deep learning. Finally, we discuss recent trends in MRI, open problems, and future directions for the field.
   • Link to Publisher
   • BibTeX

   @article{Wen2020-Trans-Learn,
     author = {Wen, Bihan and Ravishankar, Saiprasad and Pfister, Luke and Bresler, Yoram},
     title = {Transform Learning for Magnetic Resonance Image
                       Reconstruction: From Model-Based Learning To Building Neural
                       Networks},
     journal = {IEEE Signal Processing Magazine},
     volume = {37},
     number = {1},
     pages = {41-53},
     year = {2020},
     doi = {10.1109/msp.2019.2951469},
     url = {https://doi.org/10.1109/msp.2019.2951469}
   }
   

   
   
4. L. Pfister and Y. Bresler, “Learning Filter Bank Sparsifying Transforms,” IEEE Transactions on Signal Processing, vol. 67, no. 2, 2019.
   • Abstract
   Data is said to follow the transform (or analysis) sparsity model if it becomes sparse when acted on by a linear operator called a sparsifying transform. Several algorithms have been designed to learn such a transform directly from data, and data-adaptive sparsifying transforms have demonstrated excellent performance in signal restoration tasks. Sparsifying transforms are typically learned using small sub-regions of data called patches, but these algorithms often ignore redundant information shared between neighboring patches. We show that many existing transform and analysis sparse representations can be viewed as filter banks, thus linking the local properties of patch-based model to the global properties of a convolutional model. We propose a new transform learning framework where the sparsifying transform is an undecimated perfect reconstruction filter bank. Unlike previous transform learning algorithms, the filter length can be chosen independently of the number of filter bank channels. Numerical results indicate filter bank sparsifying transforms outperform existing patch-based transform learning for image denoising while benefiting from additional flexibility in the design process.
   • Download
   • Link to Publisher
   • BibTeX

   @article{Pfister2018-Learn-Filter,
     author = {Pfister, Luke and Bresler, Yoram},
     title = {Learning Filter Bank Sparsifying Transforms},
     journal = {IEEE Transactions on Signal Processing},
     volume = {67},
     number = {2},
     pages = {504-519},
     year = {2019},
     doi = {10.1109/tsp.2018.2883021},
     url = {https://doi.org/10.1109/tsp.2018.2883021}
   }
   

   
   
5. L. Pfister and Y. Bresler, “Bounding Multivariate Trigonometric Polynomials,” IEEE Transactions on Signal Processing, vol. 67, no. 3, 2018.
   • Abstract
   The extremal values of multivariate trigonometric polynomials are of interest in fields ranging from control theory to filter design, but finding the extremal values of such a polynomial is generally NP-Hard. In this paper, we develop simple and efficiently computable estimates of the extremal values of a multivariate trigonometric polynomial directly from its samples. We provide an upper bound on the modulus of a complex trigonometric polynomial, and develop upper and lower bounds for real trigonometric polynomials. For a univariate polynomial, these bounds are tighter than existing bounds, and the extension to multivariate polynomials is new. As an application, the lower bound provides a sufficient condition to certify global positivity of a real trigonometric polynomial.
   • Download
   • Link to Publisher
   • BibTeX

   @article{Pfister2018-Bound-Multiv,
     author = {Pfister, Luke and Bresler, Yoram},
     title = {Bounding Multivariate Trigonometric Polynomials},
     journal = {IEEE Transactions on Signal Processing},
     volume = {67},
     number = {3},
     pages = {700-707},
     year = {2018},
     doi = {10.1109/tsp.2018.2883925},
     url = {https://doi.org/10.1109/tsp.2018.2883925}
   }
   

   
   

Conference Papers

1. S. Mittal, A. K. Balla, L. Pfister, C. Stoean, K. Falahkheirkhah, and R. Bhargava, “Tumor Identification and Grading on Histopathology Images Using Deep Learning,” in United States and Canadian Academy of Pathology Annual Meeting, 2019.
   • BibTeX

   @inproceedings{Mittal2019-Tumor-Ident,
     author = {Mittal, Shachi and Balla, Andre K. and Pfister, Luke and Stoean, Catalin and Falahkheirkhah, Kianoush and Bhargava, Rohit},
     title = {Tumor Identification and Grading on Histopathology Images
                       Using Deep Learning},
     booktitle = {United States and Canadian Academy of Pathology Annual
                       Meeting},
     year = {2019},
     pages = {nil},
     month = mar,
     month_numeric = {3}
   }
   

   
   
2. I. Kang, A. Grant, M. Dinu, J. Jaques, L. Pfister, R. Bhargava, and S. Carney, “Agile optoelectronic fiber sources for hyperspectral chemical sensing from SWIR to LWIR,” in Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XXIV, 2018.
   • Abstract
   The use of optical frequency combs (OFCs) for multi-heterodyne spectroscopy has enabled unprecedented measurement capabilities for spectroscopic sensing, including rapid acquisition speed, high resolution and high sensitivity1,2. Development of field deployable OFC sources that are widely tunable in the important chemical fingerprint region in the long-wavelength infrared (LWIR) is a major research challenge. In this paper, we report our recent efforts towards developing LWIR comb source for SILMARILS (Standoff ILluminator for Measuring Absorbance and Reflectance Infrared Light Signatures) program by IARPA. LGS has developed fiber optic sources producing spectral combs in the SWIR (1.52 to 1.56 μm and 1.7 to 2.0 μm) and in the LWIR (7.7 to 12.1 μm) regions. The spectral combs in the LWIR are generated by difference-frequency mixing one OFC centered around 1.54 μm with another OFC, whose center wavelength is tunable between 1.7 and 2.0 μm, in a nonlinear optical crystal. Average power of the generated LWIR is 1.2-12 mW and its instantaneous spectral breadth of the combs is > 80 cm-1, sufficiently broad to cover multiple molecular absorption peaks. We demonstrate standoff sensing of chemical targets having concentration as low as 12 μg/cm2 by measuring LWIR transflectance spectra using the comb source.
   • Link to Publisher
   • BibTeX

   @inproceedings{Kang2018,
     author = {Kang, Inuk and Grant, Andrew and Dinu, Mihaela and Jaques, James and Pfister, Luke and Bhargava, Rohit and Carney, Scott},
     title = {Agile optoelectronic fiber sources for hyperspectral chemical sensing from {SWIR} to {LWIR}},
     booktitle = {Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XXIV},
     year = {2018},
     pages = {nil},
     month = may,
     doi = {10.1117/12.2305120},
     month_numeric = {5}
   }
   

   
   
3. B. Wen, Y. Li, L. Pfister, and Y. Bresler, “Joint Adaptive Sparsity and Low-Rankness on the Fly: An Online Tensor Reconstruction Scheme for Video Denoising,” in 2017 IEEE International Conference on Computer Vision (ICCV), 2017.
   • Abstract
   Recent works on adaptive sparse and low-rank signal modeling have demonstrated their usefulness, especially in image/video processing applications. While a patch-based sparse model imposes local structure, low-rankness of the grouped patches exploits non-local correlation. Applying either approach alone usually limits performance in various low-level vision tasks. In this work, we propose a novel video denoising method, based on an online tensor reconstruction scheme with a joint adaptive sparse and low-rank model, dubbed SALT. An efficient and unsupervised online unitary sparsifying transform learning method is introduced to impose adaptive sparsity on the fly. We develop an efficient 3D spatio-temporal data reconstruction framework based on the proposed online learning method, which exhibits low latency and can potentially handle streaming videos. To the best of our knowledge, this is the first work that combines adaptive sparsity and low-rankness for video denoising, and the first work of solving the proposed problem in an online fashion. We demonstrate video denoising results over commonly used videos from public datasets. Numerical experiments show that the proposed video denoising method outperforms competing methods.
   • Download
   • Link to Publisher
   • BibTeX

   @inproceedings{Wen2017,
     author = {Wen, Bihan and Li, Yanjun and Pfister, Luke and Bresler, Yoram},
     title = {Joint Adaptive Sparsity and Low-Rankness on the Fly: An Online Tensor Reconstruction Scheme for Video Denoising},
     booktitle = {2017 IEEE International Conference on Computer Vision (ICCV)},
     year = {2017},
     pages = {nil},
     month = oct,
     doi = {10.1109/iccv.2017.35},
     month_numeric = {10}
   }
   

   
   
4. L. Pfister and Y. Bresler, “Automatic parameter tuning for image denoising with learned sparsifying transforms,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017.
   • Abstract
   Data-driven and learning-based sparse signal models outperform analytical models (e.g, wavelets), for image denoising, but require careful parameter tuning to reach peak performance. In this work, we provide a solution to the problem of parameter tuning for image denoising with transform sparsity regularization. We show that by viewing a learned sparsifying transform as a filter bank we can utilize the SURELET denoising algorithm to automatically tune parameters for an image denoising task. Numerical experiments show that combining SURELET with a learned sparsifying transform provides the best of both worlds. Our approach requires no parameter tuning for image denoising, yet outperforms SURELET with analytic transforms and matches the performance of transform learning denoising with hand-tuned parameters.
   • Download
   • Slides
   • BibTeX

   @inproceedings{Pfister2017,
     title = {Automatic parameter tuning for image denoising with learned sparsifying transforms},
     author = {Pfister, Luke and Bresler, Yoram},
     booktitle = {2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
     year = {2017}
   }
   

   
   
5. L. Pfister, Y. Bresler, R. Bhargava, and P. S. Carney, “Inverse Scattering with Chemical Composition Constraints for Spectroscopic Tomography,” in Imaging and Applied Optics, 2016.
   • Abstract
   A dramatic reduction in data required for chemically specific 3-D imaging is achieved through prior constraints on the known constituents of the sample. We solve the inverse scattering problem to determine morphology and composition.
   • Download
   • Link to Publisher
   • BibTeX

   @inproceedings{Pfister2016b,
     author = {Pfister, Luke and Bresler, Yoram and Bhargava, Rohit and Carney, P. Scott},
     title = {Inverse Scattering with Chemical Composition Constraints for Spectroscopic Tomography},
     booktitle = {Imaging and Applied Optics},
     year = {2016},
     pages = {nil},
     month = jul,
     publisher = {Optical Society of America},
     doi = {10.1364/math.2016.mw2i.3},
     journal = {Imaging and Applied Optics 2016},
     month_numeric = {7}
   }
   

   
   
6. L. Pfister and Y. Bresler, “Learning Sparsifying Filter Banks,” in Proc. SPIE Wavelets & Sparsity XVI, 2015, vol. 9597.
   • Abstract
   Recent years have numerous algorithms to learn a sparse synthesis or analysis model from data. Recently, a generalized analysis model called the ’transform model’ has been proposed. Data following the transform model is approximately sparsified when acted on by a linear operator called a sparsifying transform. While existing transform learning algorithms can learn a transform for any vectorized data, they are most often used to learn a model for overlapping image patches. However, these approaches do not exploit the redundant nature of this data and scale poorly with the dimensionality of the data and size of patches. We propose a new sparsifying transform learning framework where the transform acts on entire images rather than on patches. We illustrate the connection between existing patch-based transform learning approaches and the theory of block transforms, then develop a new transform learning framework where the transforms have the structure of an undecimated filter bank with short filters. Unlike previous work on transform learning, the filter length can be chosen independently of the number of filter bank channels. We apply our framework to accelerating magnetic resonance imaging. We simultaneously learn a sparsifying filter bank while reconstructing an image from undersampled Fourier measurements. Numerical experiments show our new model yields higher quality images than previous patch based sparsifying transform approaches.
   • Download
   • Slides
   • Link to Publisher
   • BibTeX

   @inproceedings{Pfister2015,
     title = {Learning Sparsifying Filter Banks},
     author = {Pfister, Luke and Bresler, Yoram},
     booktitle = {Proc. SPIE Wavelets \& Sparsity XVI},
     year = {2015},
     month = aug,
     publisher = {SPIE},
     volume = {9597},
     doi = {10.1117/12.2188663},
     month_numeric = {8}
   }
   

   
   
7. L. Pfister and Y. Bresler, “Tomographic reconstruction with adaptive sparsifying transforms,” in 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2014.
   • Abstract
   A central problem in computed tomography (CT) imaging is to obtain useful, high-quality images from low-dose measurements. Methods that exploit the sparse representations of tomographic images have long been known to improve the quality of reconstructions from low-dose data. Recent work has shown that sparse representations learned directly from the data can outperform traditional, fixed representations, but are prohibitively expensive for practical use in CT. We propose a new method for tomographic reconstruction from low-dose data by combining the statistically weighted data fidelity term with an adaptive sparsifying transform regularizer. This regularizer can be fit to the data at lower cost than competing methods. Our algorithm alternates between reconstructing the image and learning the sparsifying transform. The Alternating Direction Method of Multipliers technique is used to provide an efficient solution to the statistically weighted minimization problem. Numerical experiments on data from clinical CT reconstructions indicate that adaptive sparsifying transform regularization outperforms synthesis sparsity methods at speeds rivaling total-variation regularization.
   • Download
   • Slides
   • Link to Publisher
   • BibTeX

   @inproceedings{Pfister2014a,
     title = {Tomographic reconstruction with adaptive sparsifying transforms},
     author = {Pfister, Luke. and Bresler, Yoram.},
     booktitle = {2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
     year = {2014},
     month = may,
     pages = {6914--6918},
     doi = {10.1109/ICASSP.2014.6854940},
     month_numeric = {5}
   }
   

   
   
8. L. Pfister and Y. Bresler, “Model-based iterative tomographic reconstruction with adaptive sparsifying transforms,” in Proc. SPIE Computational Imaging XII, 2014.
   • Abstract
   Model based iterative reconstruction algorithms are capable of reconstructing high-quality images from lowdose CT measurements. The performance of these algorithms is dependent on the ability of a signal model to characterize signals of interest. Recent work has shown the promise of signal models that are learned directly from data. We propose a new method for low-dose tomographic reconstruction by combining adaptive sparsifying transform regularization within a statistically weighted constrained optimization problem. The new formulation removes the need to tune a regularization parameter. We propose an algorithm to solve this optimization problem, based on the Alternating Direction Method of Multipliers and FISTA proximal gradient algorithm. Numerical experiments on the FORBILD head phantom illustrate the utility of the new formulation and show that adaptive sparsifying transform regularization outperforms competing dictionary learning methods at speeds rivaling total-variation regularization.
   • Download
   • Slides
   • Link to Publisher
   • BibTeX

   @inproceedings{Pfister2014,
     title = {Model-based iterative tomographic reconstruction with adaptive sparsifying transforms},
     author = {Pfister, Luke and Bresler, Yoram},
     booktitle = {Proc. SPIE Computational Imaging XII},
     year = {2014},
     editor = {Bouman, Charles A. and Sauer, Ken D},
     month = mar,
     publisher = {SPIE},
     doi = {10.1117/12.2041011},
     month_numeric = {3}
   }
   

   
   
9. L. Pfister and Y. Bresler, “Adaptive Sparsifying Transforms for Iterative Tomographic Reconstruction,” in International Conference on Image Formation in X-Ray Computed Tomography, 2014.
   • Abstract
   A major challenge in computed tomography imaging is to obtain high-quality images from low-dose measurements. Key to this goal are computationally efficient reconstruction algorithms combined with detailed signal models. We show that the recently introduced adaptive sparsifying transform (AST) signal model provides superior reconstructions from low-dose data at significantly lower cost than competing dictionary learning methods. We further accelerate this technique for tomography by utilizing the Linearized Alternating Direction Method of Multipliers (L-ADMM) to remove the need to solve an expensive least-squares problem that requires computing multiple forward and backward projections. Numerical experiments on data from clinical CT images show that adaptive sparsifying transform regularization outperforms total-variation and dictionary learning methods, and combining our regularizer with L-ADMM provides for faster reconstructions than standard ADMM.
   • Download
   • Slides
   • Link to Publisher
   • BibTeX

   @inproceedings{Pfister2014b,
     title = {Adaptive Sparsifying Transforms for Iterative Tomographic Reconstruction},
     author = {Pfister, Luke and Bresler, Yoram},
     booktitle = {International Conference on Image Formation in X-Ray Computed Tomography},
     year = {2014},
     pages = {107 -- 110},
     url = {http://www.ucair.med.utah.edu/CTmeeting}
   }
   

   
   

PhD Thesis

Masters Thesis

1. L. Pfister, “Tomographic Reconstruction with Adaptive Sparsifying Transforms,” Master's thesis, University of Illinois at Urbana-Champaign, 2013.
   • Abstract
   A major obstacle in computed tomography (CT) is the reduction of harmful x-ray dose while maintaining the quality of reconstructed images. Methods which exploit the sparse representations of tomographic images have long been known to improve the quality of reconstructions from low-dose data. Recent work has shown the promise of adaptive, rather than fixed, sparse representations. In particular, the synthesis dictionary learning framework has been shown to outperform traditional regularization techniques. However, these methods scale poorly with data size, and may be prohibitively expensive for practical tomographic reconstruction. In this thesis, we propose a new method for image reconstruction from low-dose data. The method combines a statistical iterative reconstruction framework with an adaptive sparsifying transform penalty. An alternating minimization approach is used to jointly reconstruct the image while learning a sparsifying transform adapted to the particular image being reconstructed. The Alternating Direction Method of Multipliers is used to provide a computationally efficient solution to the statistically weighted minimization problem. Numerical experiments are performed on phantom data and clinical CT images. Dose reduction is achieved through reduction in the number of views and reduction in the photon flux. The results indicate the adaptive sparsifying transform regularization outperforms state-of-the-art synthesis sparsity methods at speeds rivaling total-variation regularization.
   • Link to Publisher
   • BibTeX

   @mastersthesis{Pfister2013,
     title = {Tomographic Reconstruction with Adaptive Sparsifying Transforms},
     author = {Pfister, Luke},
     school = {University of Illinois at Urbana-Champaign},
     year = {2013},
     url = {http://hdl.handle.net/2142/45347}
   }