August 17, 2026 Dataset Open Access

Starke, Sebastian; Smid, Michal

DCAT Export

<?xml version='1.0' encoding='utf-8'?>
<rdf:RDF xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#" xmlns:adms="http://www.w3.org/ns/adms#" xmlns:cnt="http://www.w3.org/2011/content#" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:dct="http://purl.org/dc/terms/" xmlns:dctype="http://purl.org/dc/dcmitype/" xmlns:dcat="http://www.w3.org/ns/dcat#" xmlns:duv="http://www.w3.org/ns/duv#" xmlns:foaf="http://xmlns.com/foaf/0.1/" xmlns:frapo="http://purl.org/cerif/frapo/" xmlns:geo="http://www.w3.org/2003/01/geo/wgs84_pos#" xmlns:gsp="http://www.opengis.net/ont/geosparql#" xmlns:locn="http://www.w3.org/ns/locn#" xmlns:org="http://www.w3.org/ns/org#" xmlns:owl="http://www.w3.org/2002/07/owl#" xmlns:prov="http://www.w3.org/ns/prov#" xmlns:rdfs="http://www.w3.org/2000/01/rdf-schema#" xmlns:schema="http://schema.org/" xmlns:skos="http://www.w3.org/2004/02/skos/core#" xmlns:vcard="http://www.w3.org/2006/vcard/ns#" xmlns:wdrs="http://www.w3.org/2007/05/powder-s#">
  <rdf:Description rdf:about="https://doi.org/10.14278/rodare.4956">
    <rdf:type rdf:resource="http://www.w3.org/ns/dcat#Dataset"/>
    <dct:type rdf:resource="http://purl.org/dc/dcmitype/Dataset"/>
    <dct:identifier rdf:datatype="http://www.w3.org/2001/XMLSchema#anyURI">https://doi.org/10.14278/rodare.4956</dct:identifier>
    <foaf:page rdf:resource="https://doi.org/10.14278/rodare.4956"/>
    <dct:creator>
      <rdf:Description rdf:about="http://orcid.org/0000-0001-5007-1868">
        <rdf:type rdf:resource="http://xmlns.com/foaf/0.1/Agent"/>
        <foaf:name>Starke, Sebastian</foaf:name>
        <foaf:givenName>Sebastian</foaf:givenName>
        <foaf:familyName>Starke</foaf:familyName>
        <org:memberOf>
          <foaf:Organization>
            <foaf:name>HZDR</foaf:name>
          </foaf:Organization>
        </org:memberOf>
      </rdf:Description>
    </dct:creator>
    <dct:creator>
      <rdf:Description rdf:about="http://orcid.org/0000-0002-7162-7500">
        <rdf:type rdf:resource="http://xmlns.com/foaf/0.1/Agent"/>
        <foaf:name>Smid, Michal</foaf:name>
        <foaf:givenName>Michal</foaf:givenName>
        <foaf:familyName>Smid</foaf:familyName>
        <org:memberOf>
          <foaf:Organization>
            <foaf:name>HZDR</foaf:name>
          </foaf:Organization>
        </org:memberOf>
      </rdf:Description>
    </dct:creator>
    <dct:title>Software: removal of bremsstrahlung background from SAXS signals with deep neural networks</dct:title>
    <dct:publisher>
      <foaf:Agent>
        <foaf:name>Rodare</foaf:name>
      </foaf:Agent>
    </dct:publisher>
    <dct:issued rdf:datatype="http://www.w3.org/2001/XMLSchema#gYear">2026</dct:issued>
    <dcat:keyword>SAXS</dcat:keyword>
    <dcat:keyword>XFEL</dcat:keyword>
    <dcat:keyword>equivariant neural networks</dcat:keyword>
    <dcat:keyword>noise removal</dcat:keyword>
    <dct:issued rdf:datatype="http://www.w3.org/2001/XMLSchema#date">2026-08-17</dct:issued>
    <dct:language rdf:resource="http://publications.europa.eu/resource/authority/language/ENG"/>
    <owl:sameAs rdf:resource="https://rodare.hzdr.de/record/4956"/>
    <adms:identifier>
      <adms:Identifier>
        <skos:notation rdf:datatype="http://www.w3.org/2001/XMLSchema#anyURI">https://rodare.hzdr.de/record/4956</skos:notation>
      </adms:Identifier>
    </adms:identifier>
    <owl:sameAs rdf:resource="https://www.hzdr.de/publications/Publ-37977"/>
    <dct:isVersionOf rdf:resource="https://doi.org/10.14278/rodare.2585"/>
    <dct:isPartOf rdf:resource="https://rodare.hzdr.de/communities/rodare"/>
    <owl:versionInfo>1</owl:versionInfo>
    <dct:description>&lt;p&gt;This record contains the trained pytorch-based deep learning network for removing bremsstrahlung background from SAXS imaging data obtained at the European XFEL laboratory.&lt;/p&gt; &lt;p&gt;We thank Peter Steinbach for providing the codebase for the equivariant UNet, which we used as a backbone for our network. The training code is available at&amp;nbsp;https://github.com/Helmholtz-AI-Matter/bremsstrahlung_denoising.&lt;/p&gt; &lt;p&gt;Below we share a brief description of our method:&lt;/p&gt; &lt;ol&gt; &lt;li&gt;&lt;strong&gt;Introduction&lt;/strong&gt; &lt;p&gt;Experimental data from cameras in ultra-high intensity laser interaction experiments very often con-&lt;br&gt; tains not only the desired signal, but also a large amount of traces of high-energy photons created&lt;br&gt; via the bremsstrahlung process during the interaction. For example, the Jungfrau camera detecting&lt;br&gt; small angle x-ray scattering (SAXS) signal in a combined XFEL + optical laser (OL) experiment at&lt;br&gt; the European XFEL laboratory still contains lot of bremsstrahlung background, even though strong&lt;br&gt; experimental effort (adding a mirror to reflect the signal, and a massive lead wall to block direct view)&lt;br&gt; was taken to reduce those (&amp;Scaron;mı́d et al., 2020). Especially in the SAXS case, the signal is gradually&lt;br&gt; becoming weaker with increasing scattering angle. Therefore, the experimentally observed signal-to-&lt;br&gt; noise ratio determines the limit of the scattering angles for which the signal can be extracted, limiting&lt;br&gt; the physics that can be observed.&lt;br&gt; As the noise is produced by the high-energy photons, whose origin is very different from the signal&lt;br&gt; photons, the signal and noise are additive. The currently used Jungfrau camera has a resolution of&lt;br&gt; 1024 &amp;times; 512 pixels, pixel size of 75 &amp;mu;m, and the read values are calibrated to deposited keV per pixel.&lt;/p&gt; &lt;/li&gt; &lt;li&gt;&lt;strong&gt;Methods&lt;/strong&gt;&lt;br&gt; The process of removing the noise from the data was split into three steps. First, the learning dataset&lt;br&gt; was curated and cut into patches of 128 &amp;times; 128 pixels. Second, a neural network was created an trained&lt;br&gt; on those data. Splitting the data into the patches actually enables the whole process, because no&lt;br&gt; &amp;lsquo;noise-only&amp;rsquo; data are measured in the detector areas where signal typically is. In the third step, an&lt;br&gt; image with actual data is split into the patches, those are processed by the neural network, and merged&lt;br&gt; together to produce the final signal and noise prediction.&lt;br&gt; &lt;br&gt; &lt;strong&gt;Data preparation&lt;/strong&gt;&lt;br&gt; The experimental data used for training the neural network came from two sets:&lt;br&gt; &lt;br&gt; &amp;bull; X-ray only shots: Those data are collected when only the XFEL beam was used, i.e. they do&lt;br&gt; contain an example of the useful signal, but no bremsstrahlung background at all.&lt;br&gt; &amp;bull; Full shots: Those data are from the real physics shots, contain both the XFEL and OL beams,&lt;br&gt; therefore have a mixture of signal and noise.&lt;br&gt; &lt;br&gt; In order to train the neural network in a supervised manner, we need to provide two sets of data: the&lt;br&gt; signal and the noise patches. The signal patches are created from the x-ray only data like this: From&lt;br&gt; each image, a set of randomly positioned and randomly oriented patches is extracted. The randomness&lt;br&gt; in rotation is important, as those training x-ray data do have significant dominant directions, which&lt;br&gt; are expected to change in the real full shots data. Next, the patches are checked and only those&lt;br&gt; which have integrated intensity above a given threshold are used, to prevent close-to-empty patches&lt;br&gt; to be used for the training. In the last step, the amplitude of the patches is randomized, to keep the&lt;br&gt; algorithm more general. Note that the dynamic range of the detector as well as the signal is large,&lt;br&gt; i.e. above approximately four orders of magnitude.&lt;br&gt; The noise patches are created from the full shots data. To avoid the regions with signal to be used,&lt;br&gt; those regions are masked out. The masking is performed automatically by using a corresponding x-ray&lt;br&gt; only image. Then, patches of given size are randomly selected from the remaining data. Note that&lt;br&gt; neither rotation nor changes of amplitude are applied, as both can contain signatures of the structure&lt;br&gt; of bremsstrahlung, which could simplify the task for the neural network.&lt;br&gt; &lt;br&gt; &lt;strong&gt;Neural network&lt;/strong&gt;&lt;br&gt; In the modelling approach we followed, noise was assumed to be additive, i.e. a noisy input signal xin&lt;br&gt; can be decomposed into noise and clean signal components n and s, respectively via the relationship&lt;br&gt; xin = n + s.&lt;br&gt; The removal of the bremsstrahlung background n was achieved with the help of a convolutional&lt;br&gt; neural network, which estimated both the noise n̂ to be subtracted from the input and the denoised&lt;br&gt; image ŝ itself. More specifically, a UNet architecture (Ronneberger et al., 2015) was adopted with&lt;br&gt; four encoder blocks using 32, 64, 128 and 256 feature maps. Each encoder block consisted of two&lt;br&gt; separate convolutional layers and ReLU nonlinearities. No batch normalization was employed. The&lt;br&gt; corresponding decoder network matched the number of filters. The decoder output produced latent&lt;br&gt; feature maps l with 16 channels.&lt;br&gt; In preliminary experiments, we have found an equivariant version of the UNet, implemented us-&lt;br&gt; ing the &amp;lsquo;escnn&amp;rsquo; library (https://github.com/QUVA-Lab/escnn) (Cesa et al., 2022), to show favorable&lt;br&gt; performance compared to the original version. It consisted of 5.88 million trainable parameters and&lt;br&gt; implemented operations to make the network equivariant to input transformations under discrete ro-&lt;br&gt; tations with angles corresponding to multiples of 90 degrees.&lt;br&gt; The input to the neural network consisted of image patches of shape 128 &amp;times; 128. The training data&lt;br&gt; comprised of 1754 signal patches and another set of 4711 noise patches.&lt;br&gt; During network training, we randomly sampled a new noise patch each time a clean signal patch&lt;br&gt; was accessed, as a means of data augmentation and to avoid overfitting. The pixelwise addition of&lt;br&gt; both patches resulted in a synthetic noisy patch which was used as model input. Both summands&lt;br&gt; were treated as labels during model training. Image intensity normalization on the raw pixel values&lt;br&gt; was performed as follows: lower and upper bounds for z-score normalization were computed as the 1&lt;br&gt; and 99.95 percentiles of the noisy patch. The lower bound was subtracted from the noisy patch and&lt;br&gt; the result was divided by the difference between upper and lower bound. Subsequently, the result was&lt;br&gt; clipped to the unit range, i.e. values below zero were set to zero and values above one were reduced to&lt;br&gt; one. The same normalization and clipping strategy using the bounds obtained from the noisy patch&lt;br&gt; were subsequently applied on the signal and the noise patch, respectively.&lt;br&gt; From the latent representation of the equivariant UNet, pixelwise noise was estimated by further&lt;br&gt; applying a convolutional layer on the latent feature map, using a kernel size of three, with stride and&lt;br&gt; padding of one to retain the spatial dimensionality. A ReLU activation was applied, as the noise&lt;br&gt; contribution was known to be non-negative. The estimated noise ŝ was then subtracted from the&lt;br&gt; input. To enforce non-negativity also of the estimated signal, again, a ReLU nonlineariy was applied.&lt;br&gt; In total, the procedure worked as follows:&lt;br&gt; l = eqUNet(xin ),&lt;br&gt; n̂ = ReLU (conv(l)) ,&lt;br&gt; ŝ = ReLU (xin &amp;minus; n̂) .&lt;br&gt; The network was implemented using the &amp;lsquo;PyTorch&amp;rsquo; library (version 1.12.1) for the Python pro-&lt;br&gt; gramming language (version 3.10.4). It was trained for 400 epochs with a batch size of 16 on a single&lt;br&gt; NVIDIA A100 GPU using the AdamW optimizer with a learning rate of 10&amp;minus;4 and no weight decay. For&lt;br&gt; both estimated components n̂ and ŝ, the mean absolute error loss was applied. Both loss components&lt;br&gt; were added to obtain the loss function the model was trained on.&lt;br&gt; &lt;br&gt; &lt;strong&gt;Application&lt;/strong&gt;&lt;br&gt; Once the model was trained, the removal of the bremsstrahlung background of full-sized experimental&lt;br&gt; imaging data was performed by applying the model on image patches, followed by a recombination&lt;br&gt; of the patch predictions to obtain full-sized model predictions. A simple sliding-window approach,&lt;br&gt; i.e. a regular splitting of image data into non-overlapping patches and consequent combination would&lt;br&gt; produce unwanted effects on the borders between patches, therefore a more complex method was&lt;br&gt; developed.&lt;br&gt; Each image is split into a grid of patches four times, with the following initial pixel offsets: [0,0],&lt;br&gt; [96,32], [32,96], [64,64]. Normalization of the patches is performed in the same way as described for the training procedure, before being processed by the network. The obtained predictions for each&lt;br&gt; patch are then rescaled to the original data range by undoing the normalization (i.e. by multiplying&lt;br&gt; the output with the difference between upper and lower bound followed by an addition of the lower&lt;br&gt; bound).&lt;br&gt; In the last step, the four predictions produced for the four offsets are combined into a final result.&lt;br&gt; Each pixel of the final image is calculated as a weighted mean of those four predictions. The weights&lt;br&gt; for the mean are calculated as&lt;br&gt; wi = 1 / ((|pi&amp;minus;m|/2) + 2)&lt;br&gt; where wi is the weight of i&amp;minus;th prediction pi , and m is the mean of all predictions for a given pixel.&lt;br&gt; This approach effectively eliminates the outliers, which are sometimes produced close to the edges of&lt;br&gt; the patches.&lt;br&gt; &amp;nbsp;&lt;/li&gt; &lt;li&gt;&lt;strong&gt;References&lt;/strong&gt;&lt;/li&gt; &lt;/ol&gt; &lt;p&gt;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp; [1] Cesa, G., Lang, L., &amp;amp; Weiler, M. (2022). A program to build e(n)-equivariant steerable CNNs. International Conference on Learning Representations. https: / /openreview.net/forum?id=WE4qe9xlnQw&lt;/p&gt; &lt;p&gt;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp; [2] Ronneberger, O., Fischer, P., &amp;amp; Brox, T. (2015). U-net: Convolutional networks for biomedical image segmentation. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 9351, 234&amp;ndash;241. https://doi.org/10.1007/978-3-319-24574-4 28&lt;/p&gt; &lt;p&gt;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp;&amp;nbsp; [3] &amp;Scaron;mı́d, M., Baehtz, C., Pelka, A., Laso Garcı́a, A., G&amp;ouml;de, S., Grenzer, J., Kluge, T., Konopkova, Z., Makita, M., Prencipe, I., Preston, T. R., R&amp;ouml;del, M., &amp;amp; Cowan, T. E. (2020). Mirror to measure small angle x-ray scattering signal in high energy density experiments. Review of Scientific Instruments, 91 (12), 123501. https://doi.org/10.1063/5.0021691&lt;/p&gt;</dct:description>
    <dct:accessRights rdf:resource="http://publications.europa.eu/resource/authority/access-right/PUBLIC"/>
    <dct:accessRights>
      <dct:RightsStatement rdf:about="info:eu-repo/semantics/openAccess">
        <rdfs:label>Open Access</rdfs:label>
      </dct:RightsStatement>
    </dct:accessRights>
    <dcat:distribution>
      <dcat:Distribution>
        <dct:rights>
          <dct:RightsStatement rdf:about="https://creativecommons.org/licenses/by/4.0/legalcode">
            <rdfs:label>Creative Commons Attribution 4.0 International</rdfs:label>
          </dct:RightsStatement>
        </dct:rights>
        <dcat:accessURL rdf:resource="https://doi.org/10.14278/rodare.4956"/>
      </dcat:Distribution>
    </dcat:distribution>
  </rdf:Description>
</rdf:RDF>