Erik Schou Dreier
Research Subjects
Erik Schou Dreier has a total of 58 co-authors. Here are the top co-authors by number of shared publications:
- Ulrik Lund Olsen
DTU Physics, Technical University of Denmark, Kongens Lyngby, Denmark
4 shared publications Jan Kehres
Tech Univ Denmark, Dept Phys, Fys Vej, DK-2800 Lyngby, Denmark
3 shared publications- Kim Steenstrup Pedersen
Department of Computer Science (DIKU) University of Copenhagen Copenhagen Denmark
3 shared publications - Mohamad Khalil
Department of Physics, Technical University of Denmark, Fysikvej 2800 Kgs Lyngby, Denmark
3 shared publications - Daniel Turecek
Czech Technical University in Prague
2 shared publications Matteo Busi
Technical University of Denmark, Department of Physics, Lyngby (Denmark)
2 shared publications
Latest Publications
- Open Access
- 16 July 2025
- Published by Wiley in Journal of Chemometrics
- Vol. 39 (8) , 1-17
Transforming Hyperspectral Images Into Chemical Maps: A Novel End‐to‐End Deep Learning Approach
Current approaches to chemical map generation from hyperspectral images are based on models such as partial least squares (PLS) regression, generating pixel‐wise predictions that do not consider spatial context and suffer from a high degree of noise. This study proposes an end‐to‐end deep learning approach using a modified version of U‐Net and a custom loss function to directly obtain chemical maps from hyperspectral images, skipping all intermediate steps required for traditional pixel‐wise analysis. The U‐Net is compared with the traditional PLS regression on a real dataset of pork belly samples with associated mean fat reference values. The U‐Net obtains a test set root mean squared error that is 7% lower than that of PLS regression on the task of mean fat prediction. At the same time, U‐Net generates fine detail chemical maps where 99.91% of the variance is spatially correlated. Conversely, only 2.37% of the variance in the PLS‐generated chemical maps is spatially correlated, indicating that each pixel‐wise prediction is largely independent of neighboring pixels. Additionally, while the PLS‐generated chemical maps contain predictions far beyond the physically possible range of 0%–100%, U‐Net learns to stay inside this range. Thus, the findings of this study indicate that U‐Net is superior to PLS for chemical map generation.
- 16 July 2025
- Published by Wiley in Journal of Chemometrics
- Vol. 39 (8) , 1-17
Open AccessTransforming Hyperspectral Images Into Chemical Maps: A Novel End‐to‐End Deep Learning Approach
Current approaches to chemical map generation from hyperspectral images are based on models such as partial least squares (PLS) regression, generating pixel‐wise predictions that do not consider spatial context and suffer from a high degree of noise. This study proposes an end‐to‐end deep learning approach using a modified version of U‐Net and a custom loss function to directly obtain chemical maps from hyperspectral images, skipping all intermediate steps required for traditional pixel‐wise analysis. The U‐Net is compared with the traditional PLS regression on a real dataset of pork belly samples with associated mean fat reference values. The U‐Net obtains a test set root mean squared error that is 7% lower than that of PLS regression on the task of mean fat prediction. At the same time, U‐Net generates fine detail chemical maps where 99.91% of the variance is spatially correlated. Conversely, only 2.37% of the variance in the PLS‐generated chemical maps is spatially correlated, indicating that each pixel‐wise prediction is largely independent of neighboring pixels. Additionally, while the PLS‐generated chemical maps contain predictions far beyond the physically possible range of 0%–100%, U‐Net learns to stay inside this range. Thus, the findings of this study indicate that U‐Net is superior to PLS for chemical map generation.
- Open Access
- 23 July 2024
- Published by The Open Journal in Journal of Open Source Software
- Vol. 9 (99) , 6533
IKPLS: Improved Kernel Partial Least Squares and Fast Cross-Validation Algorithms for Python with CPU and GPU Implementations Using NumPy and JAX
The ikpls software package provides fast and efficient tools for PLS (Partial Least Squares) modeling. This package is designed to help researchers and practitioners handle PLS modeling faster than previously possible-particularly on large datasets. The PLS implementations in ikpls use the fast IKPLS (Improved Kernel PLS) algorithms (Dayal & MacGregor, 1997), providing a substantial speedup compared to scikit-learn's (Pedregosa et al., 2011) PLS implementation, which is based on NIPALS (Nonlinear Iterative Partial Least Squares) (H. Wold, 1966). The ikpls package also offers an implementation of IKPLS combined with the fast cross-validation algorithm by O.-C. G. Engstrøm (2024), significantly accelerating cross-validation of PLS models-especially when using a large number of cross-validation splits. ikpls offers NumPy-based CPU and JAX-based CPU/GPU/TPU implementations. The JAX implementations are also differentiable, allowing seamless integration with deep learning techniques. This versatility enables users to handle diverse data dimensions efficiently.
- 23 July 2024
- Published by The Open Journal in Journal of Open Source Software
- Vol. 9 (99) , 6533
Open AccessIKPLS: Improved Kernel Partial Least Squares and Fast Cross-Validation Algorithms for Python with CPU and GPU Implementations Using NumPy and JAX
The ikpls software package provides fast and efficient tools for PLS (Partial Least Squares) modeling. This package is designed to help researchers and practitioners handle PLS modeling faster than previously possible-particularly on large datasets. The PLS implementations in ikpls use the fast IKPLS (Improved Kernel PLS) algorithms (Dayal & MacGregor, 1997), providing a substantial speedup compared to scikit-learn's (Pedregosa et al., 2011) PLS implementation, which is based on NIPALS (Nonlinear Iterative Partial Least Squares) (H. Wold, 1966). The ikpls package also offers an implementation of IKPLS combined with the fast cross-validation algorithm by O.-C. G. Engstrøm (2024), significantly accelerating cross-validation of PLS models-especially when using a large number of cross-validation splits. ikpls offers NumPy-based CPU and JAX-based CPU/GPU/TPU implementations. The JAX implementations are also differentiable, allowing seamless integration with deep learning techniques. This versatility enables users to handle diverse data dimensions efficiently.
- 18 April 2022
- Published by SAGE Publications in Journal of Near Infrared Spectroscopy
- Vol. 30 (3) , 107-121
Hyperspectral imaging for classification of bulk grain samples with deep convolutional neural networks
Near Infrared hyperspectral imaging (HSI) offers a fast and non-destructive method for seed quality assessment through combining spectroscopy and imaging. Recently, convolutional neural networks (CNN) have shown to be promising tools for red-green-blue (RGB) image or spectral cereal classification. This paper describes the design and implementation of deep CNN models capable of utilizing both the spatial and spectral dimension of HSI data simultaneously for analysis of bulk grain samples with densely packed kernels. Classification of eight grain samples, including six different wheat varieties, were used as a test case. The study shows that the CNN architecture ResNet, originally designed for RGB images, can be adapted to use the full spatio-spectral dimension of the HSI data through adding a linear down sample layer prior to the conventional ResNet architecture. Using traditional spectral pre-processing methods before passing the data to the CNN does not improve the classification accuracy of the networks, while a channel-wise image standardization improves the accuracy significantly. The modified ResNet applied to the full spatio-spectral dimension has a classification accuracy of up to 99.75 ± 0.02%, outperforming both purely spectral (86.5 ± 0.1%) and purely spatial (98.70 ± 0.01%) based methods in terms of accuracy, indicating that utilizing spatio-spectral correlation can improve sample classification, but also that grain classification is primarily solved using spatial information. The findings reported in this paper demonstrate how CNN networks can be designed to leverage spatio-spectral information in hyperspectral data. The combination of HSI and spatio-spectral CNN networks shows a possible method for fast prediction of bulk grain quality parameters where both spectral and spatial properties of the grains are important.
- 18 April 2022
- Published by SAGE Publications in Journal of Near Infrared Spectroscopy
- Vol. 30 (3) , 107-121
Hyperspectral imaging for classification of bulk grain samples with deep convolutional neural networks
Near Infrared hyperspectral imaging (HSI) offers a fast and non-destructive method for seed quality assessment through combining spectroscopy and imaging. Recently, convolutional neural networks (CNN) have shown to be promising tools for red-green-blue (RGB) image or spectral cereal classification. This paper describes the design and implementation of deep CNN models capable of utilizing both the spatial and spectral dimension of HSI data simultaneously for analysis of bulk grain samples with densely packed kernels. Classification of eight grain samples, including six different wheat varieties, were used as a test case. The study shows that the CNN architecture ResNet, originally designed for RGB images, can be adapted to use the full spatio-spectral dimension of the HSI data through adding a linear down sample layer prior to the conventional ResNet architecture. Using traditional spectral pre-processing methods before passing the data to the CNN does not improve the classification accuracy of the networks, while a channel-wise image standardization improves the accuracy significantly. The modified ResNet applied to the full spatio-spectral dimension has a classification accuracy of up to 99.75 ± 0.02%, outperforming both purely spectral (86.5 ± 0.1%) and purely spatial (98.70 ± 0.01%) based methods in terms of accuracy, indicating that utilizing spatio-spectral correlation can improve sample classification, but also that grain classification is primarily solved using spatial information. The findings reported in this paper demonstrate how CNN networks can be designed to leverage spatio-spectral information in hyperspectral data. The combination of HSI and spatio-spectral CNN networks shows a possible method for fast prediction of bulk grain quality parameters where both spectral and spatial properties of the grains are important.
- 28 December 2020
- Published by AIP Publishing in Applied Physics Letters
- Vol. 117 (26) , 264101
Tracking based, high-resolution single-shot multimodal x-ray imaging in the laboratory enabled by the sub-pixel resolution capabilities of the MÖNCH detector
The simultaneous retrieval of x-ray attenuation, phase, and scattering using multimodal imaging techniques is finding increasing use in a range of applications, from medicine to materials science. Most techniques rely on the mechanical movement of an optical element (e.g., a grating or a mask) to obtain the multimodal images. While single-shot approaches exist, they typically employ detector pixels smaller than the grating period, often with low detection efficiency, and are limited in resolution unless either the sample or the optical element is displaced in various positions and multiple frames are collected. In this paper, we replace mechanical motion with the MÖNCH detector's capability to reach sub-pixel resolutions by interpolating between neighboring pixels collecting the charge generated by a single x-ray event. This enabled us to obtain the pilot demonstration of a laboratory-based high-resolution, single-shot multimodal imaging technique capable of simultaneously retrieving attenuation, directional differential phase, and scatter images, without any mechanical movement. We show that our proof-of-concept setup enables a single-shot resolution of 19.5 and that the resulting images provide sufficient information to produce a reliable sample thickness map. Furthermore, we demonstrate that the setup is capable of producing single-shot directional scattering images, while leaving open the option to further increase the resolution by using sample dithering.- 28 December 2020
- Published by AIP Publishing in Applied Physics Letters
- Vol. 117 (26) , 264101
Tracking based, high-resolution single-shot multimodal x-ray imaging in the laboratory enabled by the sub-pixel resolution capabilities of the MÖNCH detector
The simultaneous retrieval of x-ray attenuation, phase, and scattering using multimodal imaging techniques is finding increasing use in a range of applications, from medicine to materials science. Most techniques rely on the mechanical movement of an optical element (e.g., a grating or a mask) to obtain the multimodal images. While single-shot approaches exist, they typically employ detector pixels smaller than the grating period, often with low detection efficiency, and are limited in resolution unless either the sample or the optical element is displaced in various positions and multiple frames are collected. In this paper, we replace mechanical motion with the MÖNCH detector's capability to reach sub-pixel resolutions by interpolating between neighboring pixels collecting the charge generated by a single x-ray event. This enabled us to obtain the pilot demonstration of a laboratory-based high-resolution, single-shot multimodal imaging technique capable of simultaneously retrieving attenuation, directional differential phase, and scatter images, without any mechanical movement. We show that our proof-of-concept setup enables a single-shot resolution of 19.5 and that the resulting images provide sufficient information to produce a reliable sample thickness map. Furthermore, we demonstrate that the setup is capable of producing single-shot directional scattering images, while leaving open the option to further increase the resolution by using sample dithering.- 14 February 2020
- Published by Optica Publishing Group in Optics Letters
- Vol. 45 (4) , 1021-1024
Single-shot, omni-directional x-ray scattering imaging with a laboratory source and single-photon localization
Omni-directional, ultra-small-angle x-ray scattering imaging provides a method to measure the orientation of micro-structures without having to resolve them. In this letter, we use single-photon localization with the Timepix3 chip to demonstrate, to the best of our knowledge, the first laboratory-based implementation of single-shot, omni-directional x-ray scattering imaging using the beam-tracking technique. The setup allows a fast and accurate retrieval of the scattering signal using a simple absorption mask. We suggest that our new approach may enable faster laboratory-based tensor tomography and could be used for energy-resolved x-ray scattering imaging.
- 14 February 2020
- Published by Optica Publishing Group in Optics Letters
- Vol. 45 (4) , 1021-1024
Single-shot, omni-directional x-ray scattering imaging with a laboratory source and single-photon localization
Omni-directional, ultra-small-angle x-ray scattering imaging provides a method to measure the orientation of micro-structures without having to resolve them. In this letter, we use single-photon localization with the Timepix3 chip to demonstrate, to the best of our knowledge, the first laboratory-based implementation of single-shot, omni-directional x-ray scattering imaging using the beam-tracking technique. The setup allows a fast and accurate retrieval of the scattering signal using a simple absorption mask. We suggest that our new approach may enable faster laboratory-based tensor tomography and could be used for energy-resolved x-ray scattering imaging.
- 1 January 2019
- Published by IOP Publishing in Journal of Instrumentation
- Vol. 14 (01) , C01011
Virtual subpixel approach for single-mask phase-contrast imaging using Timepix3
- 1 January 2019
- Published by IOP Publishing in Journal of Instrumentation
- Vol. 14 (01) , C01011
Virtual subpixel approach for single-mask phase-contrast imaging using Timepix3
- 26 October 2018
- Published by International Union of Crystallography (IUCr) in Journal of Synchrotron Radiation
- Vol. 25 (6) , 1650-1657
Subpixel resolution in CdTe Timepix3 pixel detectors
Timepix3 (256 × 256 pixels with a pitch of 55 µm) is a hybrid-pixel-detector readout chip that implements a data-driven architecture and is capable of simultaneous time-of-arrival (ToA) and energy (ToT: time-over-threshold) measurements. The ToA information allows the unambiguous identification of pixel clusters belonging to the same X-ray interaction, which allows for full one-by-one detection of photons. The weighted mean of the pixel clusters can be used to measure the subpixel position of an X-ray interaction. An experiment was performed at the European Synchrotron Radiation Facility in Grenoble, France, using a 5 µm × 5 µm pencil beam to scan a CdTe-ADVAPIX-Timepix3 pixel (55 µm × 55 µm) at 8 × 8 matrix positions with a step size of 5 µm. The head-on scan was carried out at four monochromatic energies: 24, 35, 70 and 120 keV. The subpixel position of every single photon in the beam was constructed using the weighted average of the charge spread of single interactions. Then the subpixel position of the total beam was found by calculating the mean position of all photons. This was carried out for all points in the 8 × 8 matrix of beam positions within a single pixel. The optimum conditions for the subpixel measurements are presented with regards to the cluster sizes and beam subpixel position, and the improvement of this technique is evaluated (using the charge sharing of each individual photon to achieve subpixel resolution) versus alternative techniques which compare the intensity ratio between pixels. The best result is achieved at 120 keV, where a beam step of 4.4 µm ± 0.86 µm was measured.
- 26 October 2018
- Published by International Union of Crystallography (IUCr) in Journal of Synchrotron Radiation
- Vol. 25 (6) , 1650-1657
Subpixel resolution in CdTe Timepix3 pixel detectors
Timepix3 (256 × 256 pixels with a pitch of 55 µm) is a hybrid-pixel-detector readout chip that implements a data-driven architecture and is capable of simultaneous time-of-arrival (ToA) and energy (ToT: time-over-threshold) measurements. The ToA information allows the unambiguous identification of pixel clusters belonging to the same X-ray interaction, which allows for full one-by-one detection of photons. The weighted mean of the pixel clusters can be used to measure the subpixel position of an X-ray interaction. An experiment was performed at the European Synchrotron Radiation Facility in Grenoble, France, using a 5 µm × 5 µm pencil beam to scan a CdTe-ADVAPIX-Timepix3 pixel (55 µm × 55 µm) at 8 × 8 matrix positions with a step size of 5 µm. The head-on scan was carried out at four monochromatic energies: 24, 35, 70 and 120 keV. The subpixel position of every single photon in the beam was constructed using the weighted average of the charge spread of single interactions. Then the subpixel position of the total beam was found by calculating the mean position of all photons. This was carried out for all points in the 8 × 8 matrix of beam positions within a single pixel. The optimum conditions for the subpixel measurements are presented with regards to the cluster sizes and beam subpixel position, and the improvement of this technique is evaluated (using the charge sharing of each individual photon to achieve subpixel resolution) versus alternative techniques which compare the intensity ratio between pixels. The best result is achieved at 120 keV, where a beam step of 4.4 µm ± 0.86 µm was measured.
- 25 May 2018
- Published by SPIE-Intl Soc Optical Eng in Optical Engineering
- Vol. 57 (05) , 054117-13
Spectral correction algorithm for multispectral CdTe x-ray detectors
Compared to the dual-energy scintillator detectors widely used today, energy-resolved photon-counting x-ray detectors show the potential to improve material identification in various radiography and tomography applications used for industrial and security purposes. However, detector effects, such as charge sharing and photon pileup, distort the measured spectra in pixelated, photon-counting detectors operating under high flux. These effects result in a significant performance degradation of the detectors when used for material identification where accurate spectral measurements are required. We have developed a semianalytical, postdata acquisition, computational algorithm that corrects the measured attenuation curve for severe spectral distortions caused by the detector. The calibration of the algorithm is based on simple attenuation measurements of commercially available materials using standard laboratory sources, enabling the algorithm to be used in any x-ray setup. The algorithm is developed for correcting spectral data acquired with the MultiX ME100 CdTe x-ray detector but could be adapted with small adjustments to other photon-counting, energy-resolved detectors with CdTe sensors. The validation of the algorithm has been done using experimental data acquired with both a standard laboratory source and synchrotron radiation. The experiments show that the algorithm is fast, reliable at x-ray flux up to 5 Mph / s / mm2 and greatly improves the accuracy of the measured spectrally resolved linear attenuation, making the algorithm useful for both security and industrial applications where photon-counting detectors are used.
- 25 May 2018
- Published by SPIE-Intl Soc Optical Eng in Optical Engineering
- Vol. 57 (05) , 054117-13
Spectral correction algorithm for multispectral CdTe x-ray detectors
Compared to the dual-energy scintillator detectors widely used today, energy-resolved photon-counting x-ray detectors show the potential to improve material identification in various radiography and tomography applications used for industrial and security purposes. However, detector effects, such as charge sharing and photon pileup, distort the measured spectra in pixelated, photon-counting detectors operating under high flux. These effects result in a significant performance degradation of the detectors when used for material identification where accurate spectral measurements are required. We have developed a semianalytical, postdata acquisition, computational algorithm that corrects the measured attenuation curve for severe spectral distortions caused by the detector. The calibration of the algorithm is based on simple attenuation measurements of commercially available materials using standard laboratory sources, enabling the algorithm to be used in any x-ray setup. The algorithm is developed for correcting spectral data acquired with the MultiX ME100 CdTe x-ray detector but could be adapted with small adjustments to other photon-counting, energy-resolved detectors with CdTe sensors. The validation of the algorithm has been done using experimental data acquired with both a standard laboratory source and synchrotron radiation. The experiments show that the algorithm is fast, reliable at x-ray flux up to 5 Mph / s / mm2 and greatly improves the accuracy of the measured spectrally resolved linear attenuation, making the algorithm useful for both security and industrial applications where photon-counting detectors are used.
- 14 March 2018
- Published by SPIE-Intl Soc Optical Eng in Optical Engineering
- Vol. 57 (03) , 037105-10
Simulation tools for scattering corrections in spectrally resolved x-ray computed tomography using McXtrace
Spectral computed tomography is an emerging imaging method that involves using recently developed energy discriminating photon-counting detectors (PCDs). This technique enables measurements at isolated high-energy ranges, in which the dominating undergoing interaction between the x-ray and the sample is the incoherent scattering. The scattered radiation causes a loss of contrast in the results, and its correction has proven to be a complex problem, due to its dependence on energy, material composition, and geometry. Monte Carlo simulations can utilize a physical model to estimate the scattering contribution to the signal, at the cost of high computational time. We present a fast Monte Carlo simulation tool, based on McXtrace, to predict the energy resolved radiation being scattered and absorbed by objects of complex shapes. We validate the tool through measurements using a CdTe single PCD (Multix ME-100) and use it for scattering correction in a simulation of a spectral CT. We found the correction to account for up to 7% relative amplification in the reconstructed linear attenuation. It is a useful tool for x-ray CT to obtain a more accurate material discrimination, especially in the high-energy range, where the incoherent scattering interactions become prevailing (>50 keV).
- 14 March 2018
- Published by SPIE-Intl Soc Optical Eng in Optical Engineering
- Vol. 57 (03) , 037105-10
Simulation tools for scattering corrections in spectrally resolved x-ray computed tomography using McXtrace
Spectral computed tomography is an emerging imaging method that involves using recently developed energy discriminating photon-counting detectors (PCDs). This technique enables measurements at isolated high-energy ranges, in which the dominating undergoing interaction between the x-ray and the sample is the incoherent scattering. The scattered radiation causes a loss of contrast in the results, and its correction has proven to be a complex problem, due to its dependence on energy, material composition, and geometry. Monte Carlo simulations can utilize a physical model to estimate the scattering contribution to the signal, at the cost of high computational time. We present a fast Monte Carlo simulation tool, based on McXtrace, to predict the energy resolved radiation being scattered and absorbed by objects of complex shapes. We validate the tool through measurements using a CdTe single PCD (Multix ME-100) and use it for scattering correction in a simulation of a spectral CT. We found the correction to account for up to 7% relative amplification in the reconstructed linear attenuation. It is a useful tool for x-ray CT to obtain a more accurate material discrimination, especially in the high-energy range, where the incoherent scattering interactions become prevailing (>50 keV).
- 17 January 2018
- Published by American Physical Society (APS) in Physical Review B
- Vol. 97 (1) , 014416
24-spin clusters in the mineral boleiteThe crystal structure of the mineral boleite contains clusters of 24 ions that have the shape of a truncated cube formed of eight trimers connected by edges. Susceptibility measurements and exact diagonalization calculations suggest that there are strong antiferromagnetic intratrimer interactions, such that effective degrees of freedom emerge on the trimers below K. Weaker intertrimer interactions lead to the formation of a singlet ground state for these effective spins at K. The clusters in boleite offer a situation similar to single molecule magnetism, accessible to both experiment and numerics, in which the interplay of quantum spins, geometric frustration, spin entanglement, and mesoscopic system size can be studied.- 17 January 2018
- Published by American Physical Society (APS) in Physical Review B
- Vol. 97 (1) , 014416
24-spin clusters in the mineral boleiteThe crystal structure of the mineral boleite contains clusters of 24 ions that have the shape of a truncated cube formed of eight trimers connected by edges. Susceptibility measurements and exact diagonalization calculations suggest that there are strong antiferromagnetic intratrimer interactions, such that effective degrees of freedom emerge on the trimers below K. Weaker intertrimer interactions lead to the formation of a singlet ground state for these effective spins at K. The clusters in boleite offer a situation similar to single molecule magnetism, accessible to both experiment and numerics, in which the interplay of quantum spins, geometric frustration, spin entanglement, and mesoscopic system size can be studied.