Deep Learning Approach for High-accuracy Electron Counting of Direct Electron Detectors at Increased Electron Dose
Year: 2023
Type: article
Abstract: For monolithic active pixel sensor (MAPS) direct detectors, electron counting is important for eliminating noise arising from the intrinsic variations in the deposition of energy by primary electrons (Landau noise). However, the performance of existing counting algorithms [1-3] is limited in two ways. One limitation is the correctness of recognizing clusters of pixels excited by a single electron, and the second is the accuracy of localizing the incident electron position within each cluster. For counting to be feasible, the read-out frame must be sufficiently sparse that the probability of two clusters overlapping is insignificant. Inspired by the work on training a CNN to reconstruct the incident position of electrons for the Timepix3 detector [4], we have developed a deep learning model to recognize electron events and compared its counting performance with a conventional algorithm, connected component labeling (CCL) [5]. The model is a marginal improvement in counting for sparse images, but, more importantly, it provides accurate detections of single electron events at higher electron density than conventional counting methods. We trained a Faster R-CNN [6] model as shown in Figure 2(a). The model was trained on ∼8000 simulated detector readout images of varying electron densities between 0.002 ∼ 0.1 e-/pixel. One set of final weights is saved after training on 200 keV data and another set is saved after training on 80 keV data. The workflow for applying electron counting model is shown in Figure 1(b). The raw image is first upsampled by a factor of 2 and divided into equal-sized patches using overlapping windows to achieve good performance and to limit computational complexity. To guide the model towards more accurate detection at different electron density levels, built-in hyperparameter tuning based on the number of electrons estimated from the total digital counts within the window is performed before predicting the bounding boxes. After downsampling the detected boxes and assigning the incident position to the peak intensity pixel within each box, the binary image patches are stitched together to produce the final counted image. Figure 2 shows the performance of the deep learning approach versus CCL for sparse images. At 200 keV, the voltage for which the detector is optimized, the MTF at Nyquist frequency is improved to 0.63 by the DL model vs. 0.60 by CCL. Also, the DL model MTF does not show the sharp drop at low frequency that occurs in the CCL counting and integrating mode, indicating that the model is less affected by backscattering electrons at large lateral distances from the initial incident positions. The MTF and lower averaged position deviation of the DL model (Figure 2c) demonstrate the high accuracy of the electron positions. The high DQE indicates that no extra noise is induced by the DL model detection. At 80 keV, similar characteristics of the DL model performance can be drawn as at 200 keV. Due to the larger interaction volume in the sensor at 80 kV, the electron events are larger than those at 200 keV. Therefore, the accuracy of electron counting would benefit more from a precise localization method than from the precise event recognition method implemented here. This explains the lack of improvement for the DL model in high-frequency MTF and position deviation compared to CCL at 80 kV. The DL model enables high-quality electron counting with minimal coincidence loss [7] over a much larger electron dose range than CCL. In Figure 3(a, b), recall measures the fraction of electrons that are located within the 1-pixel neighborhood of the ground truth position, and precision measures the fraction of counting algorithm detections that fall in the 1-pixel neighborhood. For typical sparse images for electron counting, the DL model and the CCL perform nearly identically. As the electron dose increases, the model maintains high precision and increasingly better recall. In Figure 3(d), the relative detector conversion efficiency (DCE) [8] evaluated on experimental images at increasing electron dose is in good agreement with the trend of recall from simulated data. At ∼ 0.06 e-/pixel, neither the DL model nor the CCL MTF reaches 1 at low frequencies due to lost counts, but the DL model MTF is higher at all frequencies. Higher noise power spectrum (NPS) also indicates the lower coincidence loss of the model. The DL model DQE remains as high as that at low dose in Figure 2b. Based on the robustness of the DL model at high electron dose and the fast readout speed of Celeritas XS camera, ∼ 0.06 e-/pixel represents 480 e-/pixel/s, which is at least ten times higher than the maximum counting rate of other MAPS detectors operating in on-the-fly counting mode [9-11]. The DL model will extend the functionalities of the Celeritas XS detector towards counting electrons at different electron doses, providing more flexibility for experimental setups, such as varying from low-dose TEM imaging with flat illumination to fast 4D-STEM with small probe illumination. We believe that the workflow of training a deep learning model and applying it to electron counting will be implementable and beneficial for other similar direct detectors [12]. Schematic of the deep learning approach for counting single electron events from detector readout. (a) schematic of components of the Faster R-CNN model. (b) schematic of the workflow for applying the electron counting model. MTF and DQE calculated using experimental data and position deviation calculated from simulated data. (a-c) For 200 keV, (d-e) for 80 keV. In (c) and (f), the green triangle represents the mean, the yellow line represents the median, the box represents the first quantile and third quantile. Performance of the DL model and CCL at increasing electron dose. (a-c) Recall, Precision, and F1 score calculated by evaluation on simulated data at 200keV, (d-g) relative DCE, MTF, NPS and DQE calculated on experimental data at 200keV. The NPS curves were normalized by an averaged value at high frequency.
Source: Microscopy and Microanalysis
Language: English
Subfield: Nuclear and High Energy Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4385072445