betaFeedback

WorksPhysics and Astronomy

Enhanced micrometre spatial resolution and usable photon flux for hybrid pixel detectors via deep-learning-based X-ray localization

Abstract

Micrometre spatial resolution is achieved for X-rays with the MÖNCH detector, a charge-integrating hybrid pixel detector with a 25 μm pixel pitch, by exploiting charge sharing between neighbouring pixels. The conventional analytical interpolation method, however, relies on a global mapping derived from the charge-sharing statistics and is applicable only to isolated single-photon clusters, which limits both the achievable resolution and the usable photon flux. To overcome these limitations, we trained deep learning models to localize single-photon and pile-up events based on high-fidelity simulation data. For knife-edge measurements at 12 keV and a low occupancy of 0.74% (photons per pixel and frame), where single-photon events dominate, the proposed method reaches a of the edge spread function of 1.36 μm, compared with 1.54 μm for the conventional eta interpolation. We demonstrate experimentally that resolving up to three photons per cluster raises the reconstruction efficiency from 59% to 93% at a 2.4{\times} higher occupancy of 1.81%, while preserving a of 1.39 μm.

Introduction

MÖNCH, a charge-integrating hybrid pixel detector with a 25 µm pixel pitch [1], provides a unique combination of high spatial resolution, kHz frame rate, and energy discrimination for X-ray imaging [2, 3]. The small pixel size enables micrometre spatial resolution for X-rays [4, 5, 6] by exploiting charge sharing via the eta interpolation method [7]: the charge-weighted centre η\eta of the single-photon cluster is mapped to the sub-pixel position using a single global mapping derived from the charge-sharing statistics. This approach, however, has two limitations regarding the achievable spatial resolution and the usable photon flux. Given the nature of X-ray absorption and the resulting variation in charge sharing with the photon absorption depth, a single global mapping cannot accurately describe the sub-pixel position for all events, which limits the achievable spatial resolution. Since it requires isolated single-photon clusters, the eta interpolation is applicable only to low-flux measurements, which limits the usable photon flux and thus the experimental throughput.

Diffraction-limited storage rings, such as MAX IV [8], ESRF-EBS [9] and SLS 2.0 [10], provide up to two orders of magnitude higher brilliance. Applications which require a detector with micrometre spatial resolution together with a high photon flux, such as X-ray Fourier ptychography [11], can therefore be carried out with a substantially higher experimental throughput. Meanwhile, the micrometre detector resolution of MÖNCH also benefits lab-based X-ray imaging, e.g. full-field transmission X-ray microscopy with sub-micrometre resolution [6, 12] and beam-tracking multimodal imaging [13]. Under these conditions the requirement of single-photon clusters limits the photon flux and results in long measurement times. The fraction of photons which can be reconstructed decreases rapidly with the detector occupancy, defined here as the average number of photons per pixel and frame1, as shown in Figure 1.

Image not extracted

Figure 1. Fraction of the reconstructable photons as a function of the detector occupancy from simulation and measurement, when clusters containing at most one, two, or three photons are accepted. Frames with Poisson-distributed 12 keV photons were processed with the cluster finder and energy windows of § 2.1.

At an occupancy of 1.81%, corresponding to the high-flux measurement of § 4.3, only 60% of the photons are expected to be contained in single-photon clusters, while the remaining photons have to be discarded. Pile-up events rapidly dominate at higher occupancies. Increasing the frame rate to reduce the occupancy poses significant challenges for the readout electronics and the data acquisition system. On the other hand, resolving pile-up events at the data analysis level can significantly increase the usable photon flux without any hardware modification.

Deep learning (DL) provides a natural way to overcome both limitations of the eta interpolation, namely the resolution loss caused by the single global mapping and the flux restriction imposed by the single-photon requirement. Rather than reducing a cluster to its charge-weighted centre, a neural network learns the nonlinear relation between the full charge-sharing pattern sampled by the pixel array and the photon position. It can therefore adapt to the charge cloud of each individual event, and it can be trained to localize several photons within the same pile-up cluster. Neural networks have been applied to position reconstruction in pixelated detectors for charged particles [14], electrons [15] and photons [16, 17]. For the MÖNCH detector, DL-based localization has been demonstrated to improve the position resolution for 200 keV electrons in electron microscopy [18] with respect to analytical methods.

The transfer of the DL approach to X-rays has so far been hindered by the lack of suitable training samples, since measured single-photon events whose incident positions are known with micrometre precision are difficult to obtain, let alone pile-up events. Two recent developments remove this obstacle. First, a charge transport simulation for X-rays incorporating the repulsion effect [19] provides high-fidelity simulated events with the ground truth. Second, a pixel-wise gain and nonlinearity calibration [20] improves both the energy accuracy and the uniformity of the pixel response, reducing the systematic differences between simulation and measurements.

In this work, deep learning models were developed to localize single-photon and pile-up events in the MÖNCH detector, trained on simulated events. The models were validated with knife-edge measurements at 12 keV at the DanMAX beamline of MAX IV. For single-photon events measured at the 0.74% low-flux condition, the edge spread function width σ\sigma is improved to 1.36±0.021.36 \pm0.02 μm\mu\mathrm{m} with respect to 1.54±0.021.54 \pm0.02 μm\mu\mathrm{m} of the eta interpolation. By resolving clusters containing up to three photons, the reconstruction efficiency is increased from 59% to 93% at an occupancy of 1.81% while preserving a σ\sigma of 1.39±0.021.39 \pm0.02 μm\mu\mathrm{m}. This paper is organized as follows. § 2 describes the detector, the reconstruction framework, the generation of the training samples and the network architecture. § 3 presents the localization performance obtained in simulation for single-photon and pile-up events, and § 4 the experimental validation. § 5 discusses the origin of the improvement, the quantitative comparison between simulation and measurement, the necessity of the pile-up reconstruction, the limitations and the computational cost, followed by the conclusions in § 6.

Detector and reconstruction framework

The MÖNCH03 detector and data processing pipeline

The MÖNCH03 detector [1] used in this experiment consists of a 400×400400 \times400 pixel array with a pixel pitch of 2525 μm\mu\mathrm{m}, bump-bonded to a 650650 μm\mu\mathrm{m}-thick silicon sensor. Each pixel comprises a charge-sensitive amplifier, a correlated double sampler and configurable gains, and the analog outputs of every 25×20025 \times200 pixel block are digitized by one 14-bit ADC channel. The fully analog, charge-integrating readout preserves the charge shared among neighbouring pixels. For the measurements reported here the detector was operated in the high gain mode at an exposure time of 50 µs, with the silicon sensor biased at 150 V and a cooling temperature of 20 °C, corresponding to a noise of 0.13 keV RMS per pixel.

The data processing pipeline is shown in Figure 2. Raw frames in analog-to-digital units (ADU) are first corrected for the pedestal, and are subsequently converted to energy by means of the pixel-wise lookup table described in [20]. Clusters are then formed by selecting a seed pixel with a signal exceeding five times its noise level and recursively adding neighbouring pixels with a signal above three times their noise level [21]. The multiplicity NN of a cluster, i.e. the number of photons it contains, is determined from its total energy: a cluster is assigned multiplicity NN if its energy lies within N×(12±2)N \times(12 \pm2) keV and clusters with N≤3N \le3 are considered in this work. For each cluster, a window of 3N×3N3N \times3N pixels (3 × 3, 6 × 6 and 9 × 9 for N=1,2,3N = 1, 2, 3) centred on its charge-weighted centre is extracted. Pixels in the window which do not belong to the cluster are retained, since they may still contain charge from the photon(s) of interest below the clustering threshold, unless they are assigned to another cluster, in which case they are set to zero. Clusters are passed to the deep learning models to reconstruct the incident position(s) of the photon(s) in the cluster. The sub-pixel positions are subsequently flattened to avoid grid artefacts as described in § 2.4. Eta interpolation [7] is used as the reference method for single-photon events.

Image not extracted

Figure 2. Data processing pipeline. Raw frames are pedestal-corrected and converted to energy using a pixel-wise lookup table; clusters are then extracted and passed to the deep learning models and the eta interpolation (for single-photon events), which return the incident positions of the photons in the cluster. The reconstructed sub-pixel positions are flattened before the image is accumulated, as is the case for the eta interpolation.

Training samples generated from simulation

The 12 keV X-ray samples were generated using the charge transport simulation incorporating the repulsion effect described in [19]. Charge carrier-wise simulations were carried out at 150 V bias and 307.8 K sensor temperature (measured at 20 °C cooling temperature) for 128 absorption depths, uniformly distributed over the 650 µm sensor thickness. For each event, the photon absorption depth was sampled from an exponential distribution and the corresponding drift time was calculated. The distribution of the collected charge carriers was then sampled from the generalized Gaussian distribution parameterized as a function of the drift time. A Fano factor of 0.13 was applied in order to model the statistical fluctuations of the number of charge carriers. Random sub-pixel shifts were introduced in the xx and yy coordinates to ensure the generalization of the network. A Gaussian noise of 0.13 keV RMS was added to each pixel. A total of 10 M single-photon events were generated in a 3 × 3 pixel window.

Pile-up samples were obtained by superimposing two or three random single-photon events within one cluster window, with the positions uniformly distributed over the window of 6 × 6 and 9 × 9 pixels respectively. The same Gaussian noise of 0.13 keV RMS was added to each pixel after the superposition. To ensure the consistency with the experimental data, the same clustering algorithm was applied to the superimposed events, and those events with more than one cluster were discarded. 16 M pile-up events were generated for each of the two-photon and three-photon samples. The datasets were split into 81% for training, 13% for validation and 6% for testing.

Neural network architecture and training

The network input consists of the calibrated pixel energies of the cluster and the xx and yy coordinates embedded in the cluster window [22], three channels in total. Given the translation equivariance of the convolutional layers, the coordinate embedding facilitates the learning of the absolute position within the cluster window. One network was trained for each cluster multiplicity.

Each network consists of a stack of two-dimensional convolutional layers, which extract features of the charge-sharing pattern. The following fully connected layers return the 2N2N coordinates of the NN photons in the cluster. In the two-photon and three-photon networks, a spatial attention module [23] weights each position of the feature map before the fully connected layers. Since only part of the cluster window contains the charge belonging to a given photon, this allows the contributions of the individual charge clouds to be weighted separately. The networks are compact, with 23.7 k trainable parameters for the single-photon network and 1.21 M and 1.95 M for the two-photon and three-photon networks, respectively. The detailed layer configuration is available in the repository given in the data availability statement.

The models were trained by minimizing a smooth L1 loss with β=0.01\beta= 0.01 in units of the pixel pitch. For clusters containing more than one photon the network returns a set of positions without a defined order, and the loss must therefore be invariant under permutations of the predictions. The loss is evaluated for all assignments between the NN predicted and the NN true positions, and the smallest value is used for back-propagation. The training was performed with PyTorch [24] using the Adam optimizer with an initial learning rate of 0.001 for 500 epochs on an NVIDIA RTX 4090 GPU, with a training time up to 12 h. The trained weights and the complete training configuration are available in the same repository.

Eta interpolation and flattening of the sub-pixel distribution

In the eta interpolation as the reference method, the charge-weighted centre ηx\eta_x and ηy\eta_y are first calculated from the 3×33 \times3 cluster. Since the resulting two-dimensional eta distribution is non-uniform, it is flattened by means of a Rosenblatt transform [25], which maps it onto a uniform sub-pixel distribution. The interpolated sub-pixel position is then obtained by applying bilinear interpolation using lookup tables derived from the transform.

The same Rosenblatt transform step is required for the output of the networks, for a reason which originates in the physics of the charge deposition. When the charge cloud is confined to one or two pixels due to deep absorption, the sub-pixel position is not measurable for two or one of the coordinates, respectively. By minimizing the residual, the network returns the pixel centre for those coordinates, which leads to a bright spot at the pixel centre and to a cross-shaped pattern in the sub-pixel distribution of the reconstructed positions, as illustrated in Figure 3(a). The Rosenblatt transform is derived for each dataset and applied to the network output, which restores a uniform distribution, Figure 3(b), at the cost of a negligible increase of the width of residual distribution (less than 0.01 μ\mum in simulation). All results reported in the following were obtained after flattening, for both methods.

Image not extracted

Figure 3. Sub-pixel distribution of the reconstructed positions for flat-field illuminated measurement samples, (a) as returned by the network and (b) after the Rosenblatt transform. The excess at the pixel centre and along the two axes through it in (a) originates from the events whose charge cloud is confined to one or two pixels, thus no sub-pixel information in both or in one of the two coordinates is available.

Simulation results

Single-photon localization

Two typical simulated single-photon events are shown in Figure 4, together with the ground truth and reconstructed positions. For event (a) with a shallow absorption and a larger cloud size, both methods reconstruct the position accurately. For event (b) with a deep absorption and a smaller cloud size, the position reconstructed by the eta interpolation is biased towards the pixel centre while the network returns a more accurate position.

Image not extracted

Figure 4. Typical simulated single-photon events for a shallow (a) and a deep (b) absorption depth, showing the ground truth position together with the positions reconstructed by the eta interpolation and by the network. The deep-absorption event produces a compact charge cloud for which the eta mapping is systematically biased.

Figure 5(a) shows the residual distributions obtained at 12 keV. Throughout this work the spatial resolution in simulation is quoted as the RMS of the one-dimensional residual between the reconstructed and the true incident position. An RMS of 0.78 μ\mum is obtained with the deep learning model, a 27% improvement compared with 1.07 μ\mum for the eta interpolation. The dependence of the spatial resolution on the charge propagation distance, i.e. the distance between the absorption point and the readout electrode, is shown in Figure 5(b). The spatial resolution degrades rapidly for propagation distances below approximately 250 μ\mum. Eta interpolation shows a minimum at 250 μ\mum, which is further discussed in § 5. For longer propagation distances the charge cloud is large enough to cover several pixels, and the resolution converges to 0.6 μ\mum and 0.8 μ\mum for the deep learning model and the eta interpolation, respectively. The deep learning model outperforms the eta interpolation over the full range, and the absolute improvement is largest at short propagation distances, where the assumption of a single global mapping exhibits the largest deviation. Since the photons enter the sensor from the side opposite to the pixel electrodes, only about 11% of the absorbed photons have a propagation distance below 50 μ\mum3. The overall resolution in Figure 5(a) is therefore dominated by the long-propagation regime, while the region of largest improvement gains weight at higher photon energies (§ 5).

Image not extracted

Figure 5. (a) Distribution of the residual between the reconstructed and the true incident position for the eta interpolation and for the network, for 12 keV photons. (b) Spatial resolution, given as the RMS of the residual, as a function of the charge propagation distance.

Pile-up localization

Typical simulated pile-up events containing two and three photons are shown in Figure 6, together with the ground truth positions and the predictions of the corresponding networks. Conventional clustering methods can no longer separate the photons with partially or strongly overlapping charge clouds, while the networks are able to reconstruct the positions of all photons in the cluster accurately.

Image not extracted

Figure 6. Typical simulated pile-up events with two (a, b) and three (c, d) photons in one cluster, showing the ground truth positions and the positions predicted by the corresponding networks.

The impact of the pile-up on the achievable spatial resolution is investigated using the two-photon pile-up events. Figure 7(a) shows the RMS of the residuals as a function of the distance between the two photons in the cluster: worse resolution is observed for smaller separations, as expected from the increasing overlap of the charge clouds. The resolution gradually improves with increasing separation, and a better resolution than the single-photon case is obtained for separations above 50 μ\mum. This is attributed to the larger charge clouds required by the larger separation to form pile-up events, as shown in Figure 7(b): these events with larger separations need large enough charge clouds to form a single cluster, which carry more sub-pixel information and thus allow for a better localization. The saturated resolution at large separations is 0.6 μ\mum, consistent with that of single-photon events with long propagation distances (Figure 5(b)). Since the RMS is dominated by the events with small separations, where the residuals are considerably larger, the overall two-photon resolution of 0.86 μ\mum nevertheless remains slightly worse than the single-photon one.

Image not extracted

Figure 7. (a) Resolution as the RMS of the residuals for the two-photon predictions as a function of the distance between the two photons in the cluster. (b) Average propagation distance of the two photons as a function of their separation. (c) Distribution of the two-photon pile-up events categorized by their separation distance. The resolution improves with increasing separation, and a better resolution than the single-photon case is obtained for separations above 50 μm, which is attributed to the larger charge clouds required by the larger separation to form pile-up events.

The results of all configurations are summarized in Table 1. The resolution achieved via deep learning degrades moderately from 0.78 μm0.78\,\mu\mathrm{m} to 0.86 μm0.86\,\mu\mathrm{m} and 1.09 μm1.09\,\mu\mathrm{m} for clusters containing one, two and three photons, respectively. Compared with the eta interpolation, however, the deep learning model exhibits a 27% improvement in spatial resolution for single-photon events and the capability to resolve pile-up events with a comparable resolution.

Cluster multiplicityMethodResolution (RMS)
Single-photoneta1.07 μm1.07\,\mu\mathrm{m}
Single-photonDL0.78 μm0.78\,\mu\mathrm{m}
Two-photonDL0.86 μm0.86\,\mu\mathrm{m}
Three-photonDL1.09 μm1.09\,\mu\mathrm{m}

Table 1. Summary of the spatial resolution obtained in simulation for single-photon and pile-up events at 12 keV. The resolution is given as the RMS of the residuals between the reconstructed and the true incident positions.

Experimental validation

Measurement setup and analysis procedure

To experimentally validate the deep learning method, knife-edge measurements were carried out at the DanMAX beamline of the MAX IV Laboratory. A monochromatic X-ray beam at a photon energy of 12 keV was used, with two flux levels obtained by inserting a different number of aluminium filters of 0.22 mm thickness each. The beam size of 1.6 mm in height and 1.3 mm in width provided flat-field illumination over the region of interest. A polished tungsten knife-edge mask was mounted a few millimetres upstream of the sensor and was slightly tilted by 4.9∘4.9^\circ with respect to the pixel columns. Flat-field images were acquired without the mask to correct the non-uniformity in the beam profile. 14 M and 7 M frames were acquired for the low- and high-flux conditions, respectively, ensuring around 10510^5 photons per pixel in the illuminated region.

The occupancies of the low- and high-flux conditions were determined from the flat-field images over the region of interest used for the edge spread function analysis (336 pixels), as the total energy of the clusters divided by the mean energy of a single-photon cluster, the number of pixels and the number of frames, yielding 0.74% and 1.81%, respectively. The reconstruction efficiency, i.e. the fraction of the absorbed photons which are reconstructed, was estimated as the ratio of the summed energy contained in the reconstructable clusters to the total energy of the same region of interest.

The data were processed with the pipeline shown in Figure 2. The average two-photon sample from simulation and measurement is shown in Figure 8, validating the consistency of the processing pipelines and the charge transport simulation. The sub-pixel reconstructed positions were flattened via the Rosenblatt transform. All images were reconstructed in bins of 2.5 μm\mu\mathrm{m} (one tenth of the pixel pitch) and were corrected with the subsequent flat-field measurement. The subdivision of each pixel into ten bins follows previous works with MÖNCH [5] and balances the acquisition time against the statistics per bin, which decreases with the bin area.

Image not extracted

Figure 8. Average two-photon clusters from the simulation and measurement to check the consistency of the charge transport simulation and the clustering algorithm. The projected 1D RMS of the two-photon cluster is 0.955 pixels and 0.934 pixels for the simulation and measurement, respectively.

The spatial resolution was evaluated from the edge spread function (ESF), which was obtained by projecting the reconstructed intensity onto the axis perpendicular to the edge. The ESF was fitted with an error function, from which the width σ\sigma was extracted. The binning adds a contribution of 2.5/12≈0.722.5/\sqrt{12} \approx0.72 μm in quadrature to σ\sigma, i.e. the RMS of a uniform distribution over one bin, which is included in all reported widths unless stated otherwise. The modulation transfer function (MTF) was calculated via the Fourier transform of the discrete line spread function (LSF), obtained by differentiating the ESF. The spatial frequency at which the MTF drops to 50%, denoted MTF50, is reported4 together with the MTF at the Nyquist frequency of 200 lp/mm.

Low flux: the single-photon dominated regime

At the lower occupancy of 0.74%, 80% of the photons are contained in single-photon clusters, where the eta interpolation and deep learning can be directly compared. The knife-edge image is shown in Figure 9 together with the corresponding ESF fits on the flat-field-corrected images. A width of σ=1.36±0.02 μm\sigma= 1.36 \pm0.02\ \mu\mathrm{m} is obtained with the deep learning model, a 12% improvement compared with σ=1.54±0.02 μm\sigma= 1.54 \pm0.02\ \mu\mathrm{m} for the eta interpolation. After subtracting the binning contribution in quadrature, the intrinsic widths correspond to 1.15 μm\mu\mathrm{m} and 1.36 μm\mu\mathrm{m}, respectively, i.e. a 15% improvement. The corresponding MTF50 values are 141 lp/mm and 127 lp/mm. It should be noted that the relative intensity on the illuminated side is higher than 1, indicating an excess of single-photon events close to the edge: photons landing next to the shadowed region have fewer neighbours and are therefore less likely to form a pile-up cluster. This effect and the difference between the measured widths and the widths expected from simulation are discussed in § 5.

Image not extracted

Figure 9. Knife-edge measurement at 12 keV and 0.74% occupancy, reconstructed accepting only single-photon clusters via (a) the eta interpolation and (c) the deep learning model with the region of interest indicated by the orange box. (b, d) plot the corresponding edge spread functions and error-function fits.

In this low-flux case, accepting clusters with up to two and three photons already raises the reconstruction efficiency to 95% and 98%, respectively, while the width of the ESF remains at 1.38 ±0.02 μm\pm0.02\ \mu\mathrm{m} in both cases.

High flux: the pile-up regime

At the higher occupancy of 1.81%, only 59% of the photons are contained in single-photon clusters. The images reconstructed with the eta interpolation and with the deep learning models accepting clusters with up to three photons are shown in Figure 10, together with the corresponding ESF fits. When only single-photon clusters are accepted, a bright rim is observed along the edge, since photons next to the shadowed region have fewer neighbours and are therefore less likely to be rejected as pile-up. The resulting intensity-dependent reconstruction efficiency, discussed in § 5, raises the relative intensity next to the edge to 1.24 and leads to a worse ESF fitting quality. With the eta interpolation, σ=1.48±0.02 μm\sigma= 1.48 \pm0.02\ \mu\mathrm{m} (χ2/NDF=1.6\chi^2/\mathrm{NDF} = 1.6) and an MTF50 of 132 lp/mm are obtained, compared with 1.54 ±0.02 μm\pm0.02\ \mu\mathrm{m} and 127 lp/mm at low flux. The difference is mainly attributed to the worse fitting quality, and the two results are compatible within 2 standard deviations. Applying the deep learning model to the same single-photon clusters gives σ=1.35±0.02 μm\sigma= 1.35 \pm0.02\ \mu\mathrm{m}, consistent with the low-flux result, and an MTF50 of 145 lp/mm.

Image not extracted

Figure 10. Knife-edge measurement at 12 keV and 1.81% occupancy, reconstructed accepting only single-photon clusters via (a) the eta interpolation and (c) accepting up to three-photon clusters via the deep learning model with the region of interest indicated by the orange box. (b, d) plot the corresponding edge spread functions and error-function fits.

Accepting clusters with up to three photons increases the reconstruction efficiency to 93%. The width of the edge spread function remains at 1.39 ±0.02 μm\pm0.02\ \mu\mathrm{m} with a corresponding MTF50 value of 140 lp/mm. Compared with the low-flux case, the resolution remains unchanged while extending the photon flux by a factor of 2.4 and removing the excess intensity next to the edge. Reconstruction efficiency, resolution and MTF50 values are summarized in Table 2 for the two occupancies and for the two reconstruction methods.

OccupancyMethodAccepted multiplicityReconstruction efficiencyESF width (σ\sigma)MTF50MTF(200 lp/mm)
0.74%eta≤1\leq1 photon80%1.54 ±\pm 0.02 μ\mum127 lp/mm0.170
0.74%DL≤1\leq1 photon80%1.36 ±\pm 0.02 μ\mum141 lp/mm0.225
0.74%DL≤2\leq2 photons95%1.38 ±\pm 0.02 μ\mum138 lp/mm0.221
0.74%DL≤3\leq3 photons98%1.38 ±\pm 0.02 μ\mum137 lp/mm0.229
1.81%eta≤1\leq1 photon59%1.48 ±\pm 0.02 μ\mum132 lp/mm0.187
1.81%DL≤1\leq1 photon59%1.35 ±\pm 0.02 μ\mum145 lp/mm0.229
1.81%DL≤2\leq2 photons84%1.37 ±\pm 0.02 μ\mum143 lp/mm0.223
1.81%DL≤3\leq3 photons93%1.39 ±\pm 0.02 μ\mum140 lp/mm0.244

Table 2. Summary of the knife-edge measurements at 12 keV: reconstruction efficiency, width σ\sigma of the edge spread function, MTF50 (the spatial frequency at 50% contrast) and the MTF at the Nyquist frequency of 200 lp/mm, for the two occupancies and for the different reconstruction methods. DL denotes the deep learning models. The ESF widths include the binning contribution of 0.72 μ\mum.

Discussion

Origin of the improved spatial resolution: The eta interpolation applies a single mapping to all clusters, irrespective of the absorption depth and thus of the size of the charge cloud. The optimal resolution is achieved at a propagation distance of about 250 μm\mu\mathrm{m} (Figure 5b), while too small or too large charge clouds lead to a degradation of the resolution. On the other hand, the network adapts implicitly to the individual charge cloud by extracting the features of the charge-sharing pattern of each cluster. The absolute improvement is therefore largest at short propagation distances (Figure 5b) with compact charge clouds, where the global mapping deviates most from the actual charge sharing. Since the fraction of such photons increases with the photon energy, the advantage of the deep learning models is expected to become more pronounced at higher energies: in the 650 μ\mum-thick sensor, the fraction of the absorbed photons with a propagation distance below 250 μ\mum increases from about 11% at 12 keV to about 30% at 20 keV, while it is below 1% at 8 keV.

From simulation to measurement: The models are trained exclusively on simulated events, and their agreement with the experimental data relies on the two prerequisites introduced in § 1, namely a charge transport simulation including the repulsion effect [19] and a pixel-wise gain and nonlinearity calibration [20]. The agreement between simulation and measurement is validated by the average two-photon clusters in Figure 8, which are consistent within 2% in terms of the projected RMS. The simulation thus facilitates the training by providing samples with the ground truth, which are difficult to obtain from measurements especially for pile-up events.

To compare simulation and measurement regarding the achievable resolution, the knife-edge measurements at both occupancies were also simulated at the frame level and analysed with the same reconstruction, ESF construction and fit as the measured data. The results are summarized in Table 3: the measured ESF widths are systematically larger than the simulated ones. This difference is attributed to experimental contributions such as the quality of the knife edge, the residual angle between the beam and the sensor normal. In particular, pixels carrying a signal exhibit additional dynamic noise beyond the pedestal fluctuation modelled in the simulation, which is expected to degrade the energy resolution and thus the position reconstruction.

OccupancyMethodAccepted multiplicitySimulated ESF widthExpected ESF widthMeasured ESF width
0.74%eta≤1 photon1.19 ± 0.03 µm1.54 µm†1.54 ± 0.02 µm
0.74%DL≤1 photon0.97 ± 0.02 µm1.37 ± 0.04 µm1.36 ± 0.02 µm
0.74%DL≤2 photons0.98 ± 0.03 µm1.38 ± 0.04 µm1.38 ± 0.02 µm
0.74%DL≤3 photons0.97 ± 0.02 µm1.37 ± 0.03 µm1.38 ± 0.02 µm
1.81%DL≤2 photons0.98 ± 0.01 µm1.38 ± 0.03 µm1.37 ± 0.02 µm
1.81%DL≤3 photons0.98 ± 0.01 µm1.38 ± 0.03 µm1.39 ± 0.02 µm

Table 3. Measured ESF widths compared with the widths expected from Eq. (1) with σext=0.97\sigma_{\mathrm{ext}} = 0.97 µm. The simulated widths are obtained from the frame-level simulation of the knife-edge measurements and include the 2.5 µm binning, the non-Gaussian tails of the residual distributions and, for multi-photon acceptance, the mixture of multiplicities. For the 1.81% occupancy, only the results with multi-photon acceptance are listed, since the single-photon results are affected by the bright rim discussed in § 4.3 and the resulting worse ESF fitting quality.

Assuming that these contributions add a single, method-independent term σext\sigma_{\mathrm{ext}} in quadrature, the expected ESF width is

σexp2=σsim2+σext2(1)\sigma_{\mathrm{exp}}^2 = \sigma_{\mathrm{sim}}^2 + \sigma_{\mathrm{ext}}^2 \tag*{(1)}

where σsim\sigma_{\mathrm{sim}} is the width obtained from the simulated knife-edge measurement. From the single-photon result of the eta interpolation at 0.74% occupancy, σext=0.97±0.04\sigma_{\mathrm{ext}} = 0.97 \pm0.04 μm is obtained. With this value, Eq. (1) reproduces the measured widths of all deep learning configurations within the uncertainties. Since σext\sigma_{\mathrm{ext}} is determined from the eta interpolation alone, this agreement shows that the resolution gain predicted by the simulation is fully retained in the measurement, which validates the simulation-based methodology.

Necessity of pile-up reconstruction: Even when only single-photon clusters are accepted, the number of reconstructed photons per pixel and frame is almost doubled by increasing the occupancy from 0.74% to 1.81%, which

raises the question whether a pile-up reconstruction is necessary when the dose on the sample is not a limiting factor. The rejection of pile-up clusters, however, makes the reconstruction efficiency depend on the local intensity, as photons in bright regions are more likely to form pile-up clusters than photons next to dark regions. This effect produces the bright rim along the edge in Figure 10(a), with a relative intensity of 1.24 in the corresponding edge spread function. This rim cannot be removed by the flat-field correction, since the flat-field measurement does not contain the intensity variations due to the edge. The same effect is reproduced by the frame-level simulation, as shown in Figure 11 for the 1.81% occupancy. With single-photon acceptance, the relative intensity next to the edge rises to 1.25, and it returns to 1.03 when clusters with up to three photons are accepted, in agreement with the measurement (1.24). Such a nonlinear response distorts the image contrast and becomes more pronounced at higher occupancies. Accepting clusters with up to three photons removes the rim and increases the number of reconstructed photons by a further factor of 1.55 at the same frame rate, which reduces both the measurement time and the data volume.

Image not extracted

Figure 11. Simulated knife-edge measurement at 12 keV and 1.81% occupancy, reconstructed with the deep learning models accepting only single-photon clusters (a) and clusters with up to three photons (c) with the region of interest indicated by the orange box. (b, d) plot the corresponding edge spread functions and error-function fits. The bright rim along the edge observed in Figure 10(a) is reproduced for single-photon acceptance and is removed when the pile-up clusters are reconstructed.

Limitations of the method: According to the curves of Figure 1, the three-photon hypothesis covers more than 93% of the photons up to an occupancy of 1.81%. Above this occupancy, clusters with a higher multiplicity become relevant and additional models would be required, leading to the overhead of training and maintaining multiple models. Moreover, the method cannot be pushed to arbitrarily high pile-up. To contain all NN photons of a cluster, the area of the 3N×3N3N \times3N window grows quadratically with NN, whereas the charge of each photon still covers only a few pixels, so that the number of pixels containing charge grows only linearly. The fraction of pixels containing charge therefore decreases roughly as 1/N1/N, i.e. the input becomes increasingly sparse, and a larger window is also more likely to include pixels of neighbouring clusters. Larger models are therefore required to process a sparser signal, at a rapidly decreasing inference throughput. Such a categorization of the clusters furthermore relies on the monochromaticity of the beam, so it could not be used in experiments where the beam is polychromatic, such as full-field fluorescence imaging.

Computational cost: The inference throughputs of the deep learning models are 12 M events/s, 2.2 M events/s and 0.9 M events/s for the single-photon, two-photon and three-photon models, respectively, on an NVIDIA RTX 4090 GPU. As already observed for the calibration [20], the total processing time of the offline analysis is dominated by reading the data from disk and by the cluster finding algorithm, and the additional cost of the deep learning models is therefore not significant. An implementation of the integrated pipeline on GPU or FPGA is envisaged for online reconstruction.

Conclusion and outlook

Deep learning models have been developed to reconstruct the incident positions of X-ray photons in the MÖNCH detector. Simulations show that the deep learning models outperform the eta interpolation for single-photon events, with a residual RMS of 0.78 µm compared with 1.07 µm. The improvement is understood as the ability of the network to adapt to the size of the individual charge cloud, while the eta interpolation forms a single mapping for all clusters with different absorption depths and charge cloud sizes. Dedicated models resolve clusters containing two and three photons, with a residual RMS of 0.86 µm and 1.09 µm, respectively. These results were validated with knife-edge measurements. At an occupancy of 0.74%, the width of the edge spread function is reduced from 1.54 µm via the eta interpolation to 1.36 µm via deep learning. At an occupancy of 1.81%, which is 2.4 times higher, the reconstruction of clusters containing up to three photons increases the reconstruction efficiency from 59% to 93% while the spatial resolution is preserved at 1.39 µm (1.18 µm without the binning contribution). The usable photon flux is extended by a factor of 2.4 without noticeable compromise in reconstruction efficiency and spatial resolution, equivalent to an increase in the frame rate of the detector by the same factor.

For polychromatic cases, such as full-field fluorescence imaging [26], the number of photons contained in a cluster can no longer be determined analytically from the cluster energy, which motivates a frame-based model that predicts the positions and the energies of the photons directly from the calibrated frames, bypassing the cluster finding step. Such a model, together with the integration of the reconstruction into the data acquisition pipeline of the full-size MÖNCH detector [27], is the subject of ongoing work. Finally, the methodology provides a tool to explore detector designs, where the achievable spatial resolution can be predicted and different design parameters can be optimized.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

We acknowledge MAX IV Laboratory for time on Beamline DanMAX under Proposal 20251632. Research conducted at MAX IV is supported by the Swedish Research council under contract 2018-07152, the Swedish Governmental Agency for Innovation Systems under contract 2018-04969, and Formas under contract 2019-02496. DanMAX is funded by the NUFI grant no. 4059-00009B. We also thank Mads Ry Jørgensen and Innokenty Kantor for their kind support during the beamtime.

Data availability

The training and inference code, the trained model weights, simulation data used for parameterization and massive generation, and a subset of the measurement data are available at https://doi.org/10.5281/zenodo.22807453.

References

  1. [1]M. Ramilli, A. Bergamaschi, M. Andrae, M. Brückner, S. Cartier, R. Dinapoli, E. Fröjd, D. Greiffenberg, T. Hutwelker, C. Lopez-Cuenca, D. Mezza, A. Mozzanica, M. Ruat, S. Redford, B. Schmitt, X. Shi, G. Tinti, J. Zhang, Measurements with MÖNCH, a 25 μm pixel pitch hybrid pixel detector, Journal of Instrumentation 12 (01) (2017) C01071. doi:10.1088/1748-0221/12/01/C01071.
  2. [2]S. Cartier, A. Bergamaschi, R. Dinapoli, D. Greiffenberg, I. Johnson, J. Jungmann, D. Mezza, A. Mozzanica, B. Schmitt, X. Shi, M. Stamponi, J. Sun, G. Tinti, Micron resolution of MÖNCH and GOTTHARD, small pitch charge integrating detectors with single photon sensitivity, Journal of Instrumentation 9 (05) (2014) C05027. doi:10.1088/1748-0221/9/05/C05027.
  3. [3]A. Bergamaschi, M. Andră, R. Barten, C. Borca, M. Brückner, S. Chiriotti, R. Dinapoli, E. Fröjd, D. Greiffenberg, T. Huthwelker, A. Kleibert, M. Langer, M. Lebegue, C. Lopez-Cuenca, D. Mezza, A. Mozzanica, J. Raabe, S. Redford, C. Ruder, V. Scagnoli, B. Schmitt, X. Shi, U. Staub, D. Thattil, G. Tinti, C. Vaz, S. Vetter, J. Vila-Comamala, J. Zhang, The MÖNCH detector for soft X-ray, high-resolution, and energy resolved applications, Synchrotron Radiation News 31 (6) (2018) 11–15. doi:10.1080/08940886.2018.1528428.
  4. [4]S. Cartier, M. Kagias, A. Bergamaschi, Z. Wang, R. Dinapoli, A. Mozzanica, M. Ramilli, B. Schmitt, M. Brückner, E. Fröjd, D. Greiffenberg, D. Mayilyan, D. Mezza, S. Redford, C. Ruder, L. Schädler, X. Shi, D. Thattl, G. Tinti, J. Zhang, M. Stamponi, Micrometer-resolution imaging using MÖNCH: towards G₂-less grating interferometry, Journal of Synchrotron Radiation 23 (6) (2016) 1462–1473. doi:10.1107/S1600577516014788.DOI
  5. [5]S. Chiriotti, R. Barten, A. Bergamaschi, M. Brückner, M. Carulla, I. Chsherbakov, R. Dinapoli, E. Fröjd, D. Greiffenberg, S. Hasanaj, V. Hinger, T. King, P. Kozlowski, C. Lopez-Cuenca, A. Lozinskaya, F. Marone, D. Mezza, K. Moustakas, A. Mozzanica, C. Ruder, B. Schmitt, D. Thattl, O. Tolbanov, A. Tyazhev, A. Zarubin, J. Zhang, High-spatial resolution measurements with a GaAs:C sensor using the charge integrating MÖNCH detector with a pixel pitch of 25 μm, Journal of Instrumentation 17 (04) (2022) P04007. doi:10.1088/1748-0221/17/04/P04007.
  6. [6]D. Qu, W. Jiang, H.-Y. Chen, M. R. D. Veerararaj, P. Qi, A. Bergamaschi, E. Fröjd, R. Dinapoli, J. Heymes, S. Niese, C. David, J. Vila-Comamala, Pushing the limits of high-resolution laboratory full-field transmission X-ray microscopy, Optica 13 (6) (2026) 1165–1173. doi:10.1364/OPTICA.584801.DOI
  7. [7]E. Belau, R. Klanner, G. Lutz, E. Neugebauer, H. Seebrunner, A. Wylie, T. Böhringer, L. Hubbeling, P. Weihammer, J. Kemmer, U. Kötz, M. Riebesell, Charge collection in silicon strip detectors, Nuclear Instruments and Methods in Physics Research 214 (2) (1983) 253–260. doi:10.1016/0167-5087(83)90591-4.DOI
  8. [8]P. Tavares, S. C. Leemann, M. Sjöström, Å. Andersson, The MAX IV storage ring project, Journal of Synchrotron Radiation 21 (5) (2014) 862–877. doi:10.1107/S1600577514011503.DOI
  9. [9]P. Raimondi, C. Benabderrahmane, P. Berkvens, J. C. Biasci, P. Borowiec, J.-F. Bouttelle, T. Brochard, N. B. Brookes, N. Carmignani, L. R. Carver, et al., The extremely brilliant source storage ring of the european synchrotron radiation facility, Communications Physics 6 (1) (2023) 40. doi:10.1038/s42005-023-01195-z.DOI
  10. [10]A. Streun, T. Garvey, L. Rivkin, V. Schlott, T. Schmidt, P. Willmott, A. Wulich, SLS-2 – the upgrade of the Swiss Light Source, Journal of Synchrotron Radiation 25 (3) (2018) 631–641. doi:10.1107/S1600577518002722.DOI
  11. [11]K. Wakonig, A. Diaz, A. Bonnin, M. Stamponi, A. Bergamaschi, J. Ihli, M. Guizar-Sicairos, A. Menzel, X-ray Fourier ptychography, Science Advances 5 (2) (2019) eaav0282. doi:10.1126/sciadv.aav0282.DOI
  12. [12]D. Qu, S. Zhao, P. Qi, A. Bergamaschi, M. R. D. Veeraaraj, R. Dinapoli, J. Heymes, J. Vila-Comamala, C. David, Achromatic laboratory full-field transmission X-ray microscope, Photonics Research 14 (2) (2026) 392–400. doi:10.1364/PRJ.573807.
  13. [13]E. S. Dreier, A. Bergamaschi, G. K. Kallon, R. Brönnimann, U. L. Olsen, A. Olivo, M. Endrizzi, Tracking based, high-resolution single-shot multimodal X-ray imaging in the laboratory enabled by the sub-pixel resolution capabilities of the MÖNCH detector, Applied Physics Letters 117 (26) (2020) 264101. doi:10.1063/5.0027763.
  14. [14]ATLAS Collaboration, A neural network clustering algorithm for the ATLAS silicon pixel detector, Journal of Instrumentation 9 (09) (2014) P09009. doi:10.1088/1748-0221/9/09/P09009.DOI
  15. [15]J. P. van Schayck, E. van Genderen, E. Maddox, L. Roussel, H. Boulanger, E. Fröjd, J.-P. Abrahams, P. J. Peters, R. B. Ravelli, Sub-pixel electron detection using a convolutional neural network, Ultramicroscopy 218 (2020) 113091. doi:doi.org/10.1016/j.ultramic.2020.113091. URL https://www.sciencedirect.com/science/article/pii/S0304399120302424DOI
  16. [16]C. Yang, E. M. Zannoni, L.-J. Meng, Joint estimation of interaction position and energy deposition in semiconductor SPECT imaging sensors using fully connected neural network, Physics in Medicine & Biology 68 (9) (2023) 095009. doi:10.1088/1361-6560/aca740. URL https://doi.org/10.1088/1361-6560/aca740
  17. [17]A. Peirson, R. Romaini, H. Marshall, J. Steiner, L. Baldini, Deep ensemble analysis for imaging X-ray polarimetry, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 986 (2021) 164740. doi:10.1016/j.nima.2020.164740. URL https://www.sciencedirect.com/science/article/pii/S0168900220311372DOI
  18. [18]X. Xie, L. Barba Flores, B. Bejar Haro, A. Bergamaschi, E. Fröjdh, E. Müller, K. Paton, E. Poghosyan, C. Remlinger, Enhancing spatial resolution in MÖNCH for electron microscopy via deep learning, Journal of Instrumentation 19 (01) (2024) C01020. doi:10.1088/1748-0221/19/01/C01020.
  19. [19]X. Xie, A. Bergamaschi, M. Brückner, M. Carulla, R. Dinapoli, S. Ebner, K. Ferjaoui, V. Gautam, D. Greiffenberg, S. Hasanaj, J. Heymes, V. Hinger, V. Kedych, T. King, S. Li, C. Lopez-Cuenca, A. Mazzoleni, D. Mezza, K. Moustakas, A. Mozzanica, M. Müller, J. Mulvey, K. Paton, C. Ruder, B. Schmitt, P. Sieberer, S. Silletta, D. Thattil, J. Zhang, Optimization and validation of charge transport simulation for hybrid pixel detectors incorporating the repulsion effect, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 1081 (2026) 170894. doi:10.1016/j.nima.2025.170894.
  20. [20]X. Xie, A. Bergamaschi, M. Brückner, M. Carulla, R. Dinapoli, S. Ebner, K. Ferjaoui, E. Fröjdh, V. Gautam, D. Greiffenberg, S. Hasanaj, J. Heymes, V. Hinger, M. Hürst, V. Kedych, T. King, S. Li, C. Lopez-Cuenca, A. Mazzoleni, D. Mezza, K. Moustakas, A. Mozzanica, J. Mulvey, M. Müller, K. Paton, C. Posada Soto, C. Ruder, B. Schmitt, P. Sieberer, S. Silletta, D. Thattil, J. Zhang, Improved pixel-wise calibration for charge-integrating hybrid pixel detectors with performance validation, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 1084 (2026) 171227. doi:10.1016/j.nima.2025.171227.
  21. [21]E. Fröjdh, A. Bergamaschi, M. Brückner, M. Carulla, R. Dinapoli, S. Ebner, K. Ferjaoui, D. Gautam, D. Greiffenberg, S. Hasanaj, A. Haugdal, J. Heymes, V. Hinger, V. Kedych, T. King, S. Li, C. Lopez-Cuenca, A. Mazzoleni, D. Mezza, K. Moustakas, J. Mulvey, M. Müller, K. Paton, C. Posada Soto, C. Ruder, B. Schmitt, P. Sieberer, S. Silletta, D. Thattil, X. Xie, J. Zhang, Aare — a flexible data analysis library for hybrid pixel detectors, Journal of Instrumentation 20 (12) (2025) C12029. doi:10.1088/1748-0221/20/12/C12029.DOI
  22. [22]R. Liu, J. Lehman, P. Molino, F. P. Such, E. Frank, A. Sergeev, J. Yosinski, An intriguing failing of convolutional neural networks and the CoordConv solution, in: Advances in Neural Information Processing Systems (NeurIPS), Vol. 31, 2018, pp. 9656–9667.arxiv.org/abs/1807.03247
  23. [23]S. Woo, J. Park, J.-Y. Lee, I. S. Kweon, CBAM: Convolutional block attention module, in: Proceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 3–19.arxiv.org/abs/1807.06521
  24. [24]A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, PyTorch: an imperative style, high-performance deep learning library, in: Advances in Neural Information Processing Systems, Vol. 32, 2019.arxiv.org/abs/1912.01703
  25. [25]M. Rosenblatt, Remarks on a multivariate transformation, The Annals of Mathematical Statistics 23 (3) (1952) 470–472. doi:10.1214/aoms/1177732934.
  26. [26]L. Li, S. Zhang, R. Li, Z. Chen, Full-field fan-beam X-ray fluorescence computed tomography with a conventional X-ray tube and photon-counting detectors for fast nanoparticle bioimaging, Optical Engineering 56 (4) (2017) 043106. doi:10.1117/1.OE.56.4.043106.DOI
  27. [27]J. Heymes, Characterisation of MÖNCH 0.5, a 25 μm pitch prototype charge-integrating hybrid pixel detector, conference talk, High Precision X-Ray Measurements 2025, Laboratori Nazionali di Frascati, Italy, 20 June 2025 (2025). URL https://agenda.infn.it/event/43727/contributions/264310/

Paper details

vlnc:W_3vkpaf7q
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents