1. Introduction
Much interest in the aneutronic fusion reaction of proton-boron11 (p-B11) has risen lately due to the unexpectedly large amount of reactions obtained in several high intensity laser experiments (see, e.g., [Reference Eliezer, Hora, Martinez Val, Belloni and Nissim1, Reference Margarone, Bonvalet and Giuffrida2]).
As the physical processes involved in the production of such excess of reactions are not yet fully understood and may even involve nonlinear processes such as the avalanche process [Reference Eliezer, Hora, Korn, Nissim and Martinez Val3, Reference Belloni4], intensive research programs are ongoing in many academic institutes and private companies [Reference Hora, Eliezer and Kirchhoff5–7]. Moreover, a high flux of highly energetic particles such as protons, heavy ions, and the desired fusion product of alpha-particles accompany the harsh plasma environment that characterizes laser-initiated p-B11 fusion experiments, making the analysis of such experiments a nontrivial task. The most common diagnostics in such experiments are solid state nuclear track detectors such as CR-39 [Reference Tommasino, Zapparoli, Spiezia, Griffith and Espinosa8, Reference Cartwright, Shirk and Price9]. Although the CR-39 are very reliable for the measurement of the absolute alpha-particle flux, the analysis of the CR-39 passive detector involves a long and tedious process which includes chemical etching, a smart image analysis, and a manual revision, thus limiting the accuracy and efficiency of the data analysis. For this reason, it is highly desirable to develop automatic software tools to carry out analysis for a large number of nuclear track detectors, hence enhancing the overall outputs of future p-B11 experiments at various laser facilities. That is the main goal behind our current research, which uses deep learning to automate the classification of neutron tracks on CR-39 detectors.
For fast neutron dosimetry, a CR-39 detector is added to a standard TLD card dosimeter. The CR-39 detector is composed of an organic polymer whose chemical name is polyallyl diglycol carbonate (PADC) and its chemical formula is C12H18O7. It is suitable for neutron personal monitoring due to its high sensitivity to protons hit [Reference Massand, Kundu, Dhairyawan and Marathe10], which has been recognized some decades ago as a basic requirement for a fruitful neutron personal monitoring [Reference Massand, Kundu, Marathe and Supe11].
CR-39 detectors hold some important advantages for serving as neutron dosimeters, among them are their low fast neutron energy threshold, their insensitivity to photon and beta irradiation, their high sensitivity over a wide range of neutron energies [Reference Pitt, Scharmann and Simmer12, Reference Morelli, Vilela and Fantuzzi13], and the low influence of environmental effects [Reference Charvat14] on its response as well as low signal fading. Although other neutron detecting technologies, such as the bubble detector [Reference d’Errico, Alberts and Matzke15], may have advantages in sensitivity, angle, and energy response dependence, the CR-39 detector still remains the most versatile, easy to carry along with a TLD card, and cheapest neutron dosimeter alternative.
Nevertheless, the main drawback of CR-39 as a fast neutrons dosimeter is its ability to accurately measure high flux neutron fields due to the tedious task of counting the neutron tracks. Other challenges for the CR-39 are distinguishing between different neutron energies and distinguishing between neutrons and alpha-particles.
2. Materials and Methods
The External Dosimetry Lab (EDL) at Soreq Nuclear Research Center (SNRC) provides dosimetry services to all radiation workers around the country, most of them are monitored monthly for X-rays and gamma and beta radiation and some of them for both thermal and fast neutrons.
Since late 70’s [Reference Eisen, Eliau and Karpinowitz16] and until recently, SNRC fast neutron dosimetry system has been an in-house system. In 2018, a new personal neutron dosimetry system, LANDAUER Neutrak ® system, has been adopted at SNRC [Reference Amit, Even-Hen, Awad, Levi, Buchbinder and Datz17]. This system is designed to measure CR-39 detectors using a Zeiss microscope, which is coupled to a CCD camera and to a robotic arm that feeds the microscope’s moving tray with plastic holders one at a time, each holder having six CR-39 detectors. The CR-39 detector dimensions are 9 × 19 mm.
After the EDL receives the CR-39 detectors from the customers, the first stage is the etching stage. The detectors are inserted for 15 hours into an etching bath filled with NaOH heated to 74°C at a concentration of 5.5 mol/l in order to enlarge the tracks size. After etching is finished, the CR-39 detectors are thoroughly rinsed to remove all etching residues, to achieve optimal optical reading conditions.
Following the rinsing stage, the second stage is imaging the etched CR-39 detectors using the CCD camera. Ten images of different areas of the CR-39 detector are taken for the fast neutron counting. The images are then analyzed by Landauer’s computer vision analysis software. Afterwards, the EDL staff either adds undetected tracks (false negatives) or deletes detected tracks (false positives) using Landauer’s data review SW, according to a set of rules adopted by the scientific lab team.
We developed for the first time to our knowledge a method that detects fast neutron tracks on CR-39 detectors, which is based on deep learning, and we name it FAINE, Fast artificial intelligence neutron detection. FAINE uses an artificial neural network of type U-Net and its development consisted with three stages. The first stage is neutrons tracks manual tagging, the second stage is neural network architecture setup and training, and the last stage was evaluating the neural network performance on a test set of neutron track images over CR-39 detectors. These three stages are discussed in detail in the following subsections. The first and third stages were carried out using a designated SW written for the user interface (UI) of the deep learning model as shown in Figure 1.
Figure 1: An example for the user interface of FAINE. At the large right pane is the 1st out of 10 fields of dosimeter number 2965035 as noted in the upper left pane. At the lower left pane, the statistics of this dosimeter are presented to the user, including predicted vs. real (input) neutron tracks, the confusion matrix, and the algorithm accuracy. In the middle pane, all detected signs are presented in zoom mode, so the user can examine them if needed.
2.1. Neutron Tracks Manual Tagging
The first stage was the tagging stage, where we picked 23 dosimeters consisting of 230 images and manually tagged all neutron tracks. Then, we used classical image processing algorithms to automatically detect all the objects in the images and match them to the tagged neutron tracks. Inside the 230 images, 23,567 objects were found and 2,615 of them were valid neutron tracks and 20,952 were not. The rules for deciding which signs are valid neutron tracks were taken from the EDL’s work instruction for developing CR-39 detectors for evaluation of fast neutrons, which was adopted from Landauer’s work instruction. This instruction relies on a worldwide knowledge and experience regarding the neutron tracks morphology. For instance, neutron tracks need to be between certain size limits, they need to have both inner bright circle and outer dark crown and they need to be clear and distinct from their surrounding background. Moreover, they also need to be evenly scattered across the detector area when accumulating large enough statistics. Of course, all the above rules for classifying neutron tracks are somewhat arbitrary, and so is the task of classifying them in every fast neutron dosimetry system. The main important directives to the EDL staff to follow for this classifying task were to obey the above rules and to be as repeatable as a human can be, so the deep learning algorithm described in the next step can learn the neutron tracks features as accurately as possible.
2.2. Neural Network Architecture Setup and Training
For the neutron tracks classification, we chose a state-of-the-art convolutional artificial neural network (ANN) of type U-Net, which was first introduced in 2015 [Reference Long, Shelhamer and Darrell18] and was fine-tuned later [Reference Ronneberger, Fischer, Brox and U-Net19]. This network's name is due to both its contracting (downsampling) and expansive (upsampling) paths, which give it the u-shaped architecture (Figure 2). The contracting path mainly serves as feature extraction for the net, while the expansive path is more for the localization of objects. U-Net architecture is commonly used for semantic segmentation tasks, e.g., processes of linking each pixel in an image to a class label, in our case either “track” or “nontrack.”
Figure 2: Basic schematics of the U-Net architecture. The model input is a raw image and its output is a segmented (masked) image of the neutron tracks. The U-Net consists of a contracting path and an expansive path (encoder-decoder). The contracting path follows the typical architecture of a convolutional network while the expansive path consists of an upsampling of the feature map followed by a 2 × 2 convolution (“upconvolution”) and two 3 × 3 convolutions, each followed by a rectified linear activation function (ReLU).
In order to feed the U-Net training stage with good quality images, we used dosimeter images that were taken under different lighting and environment conditions. We then manipulated the data with preprocessing techniques using some classical image processing algorithms such as canny edge detection, dilation, erosion, fill holes, and object detection. For the training stage, we used 80% of the total 23, 567 tagged signs that were created out of the 23 dosimeters, where each image containing such a sign is of resolution of 128 × 128 pixels.
2.3. U-Net Performance Evaluation on Neutron Tracks
After the training stage, we tested the U-Net classifier over the remaining 20% of the 23,567 tagged signs. The metric we used for evaluating the classifier’s performance is accuracy. Accuracy is defined as the number of correct predictions over the number of total predictions, so in terms of a binary classification model as in our case, we have the following definition:
where TP stands for true positives and is defined as the number of correctly predicted neutron tracks (seen as the track signs inside the green squares in Figure 3), TN stands for true negatives and is defined as the number of correctly predicted signs that are not neutron tracks (seen as the track signs inside the blue squares in Figure 3), FP stands for false positives and is defined as the number of signs that are not neutron tracks, which were falsely predicted as neutron tracks (seen as the track signs inside the red squares in Figure 3), and FN stands for false negatives and is defined as the number of neutron tracks, which were falsely predicted as signs that are not neutron tracks (seen as the track signs inside the orange square in Figure 3).
Figure 3: An example of FAINE tagging signs inside a CR-39 image. Green squares indicate true positives, blue squares indicate true negatives, red squares indicate false positives, and orange squares indicate false negatives.
3. Results and Discussion
As can be seen from Table 1, and using equation (1), when choosing the segmentation threshold to be 0.4, we get an accuracy of 96.7% on our test data.
Table 1: Confusion matrix of our U-Net model. As can be seen from the definition of accuracy, the two important quantities that contribute to high accuracy are true positives (TP) and true negatives (TN).
Threshold: 0.4.
This accuracy was achieved with respect to 4,509 different tagged signs. We can achieve different true positive rate (TPR) to false positive rate (FPR) ratios by applying different classification thresholds over our models prediction. We chose to apply a classification threshold of 0.4 in order to gain a high enough TPR while still maintaining the false negative rates small enough. Of course, this choice of threshold is application-specific, and each model in any scientific field will eventually need to take this choice in order to apply a certain model.
The effectiveness of measuring fast neutron dosimeters at our EDL will dramatically improve, due to the automatic nature of our new tool, which will make the lab technician attendance redundant.
Another advantage of our new tool is the fast neutrons measurement uncertainty expected improvement. Since some of this uncertainty contribution come from the robustness uncertainty, our tool should slightly improve the overall measurement uncertainty by eliminating the worker A vs. worker B robustness term [Reference Amit, Even-Hen, Awad, Levi, Buchbinder and Datz17].
This measurement uncertainty improvement shall in turn lower the system’s detection threshold. The fast neutrons detection threshold is defined in ISO 21909 : 2015 to be “the minimum measured dose equivalent, which is significantly higher (at the 95% confidence level) than the mean dose equivalent of a sample of nonirradiated detectors.” Needless to note that the mean dose equivalent of unirradiated detectors measured by our new tool is supposed to be equal or lower than the one measured by the former method, since the former method used an overshoot dose assessment as a way to address the inherent computer vision problem to detect all fast neutron tracks.
4. Conclusions and Future Work
To conclude, we developed a novel algorithm that uses a deep learning U-Net model to accurately and repeatedly classify fast neutron tracks on CR-39 detectors with a high accuracy of 96.7%. This method can replace manual revision of track counting by an automatic repeatable process that will save a large amount of human time, especially as the number of CR-39 detectors to analyze gets higher in high neutron flux experiments.
It is worth mentioning that such deep learning methods are not limited to detection of massive particles, similar U-Net models can be also used for detection, imaging, and classification tasks with visible, X-ray and gamma photons (see, e.g., [Reference Shimobaba, Endo and Nishitsuji20, Reference Bian, Yi, Hu, Zhang, Wang and Gao21]) which are now explored at SNRC as well.
In the near future, we plan to extend our U-Net model in order to gain new capabilities for differentiating between alpha-particles and protons and for fast neutrons and alpha-particles spectrometry using machine learning classification. For this extension of our model, we already started a process of its calibration for both alpha-particles detection and for fast neutron spectrometry classification, and we expect that our model will be very useful in analyzing results of proton-boron11 fusion experiments in the future.
Data Availability
The data used to support the findings of this study are available from the corresponding author upon request.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Authors’ Contributions
Both Gal Amit and Idan Mosseri contributed equally to this research and should be considered as co-first authors.
Acknowledgments
This research was funded by the Israeli Ministry of Energy (grant number 01030961) and by the IAEC-UPBC PAZY Foundation (grant number 2472020).
Open access
