EP3977175A1

EP3977175A1 - Method and device for identifying atomic species emitting x- or gamma radiation

Info

Publication number: EP3977175A1
Application number: EP20727669.2A
Authority: EP
Inventors: Olivier Limousin; Geoffrey DANIEL; Daniel Maier
Original assignee: Commissariat a lEnergie Atomique CEA; Commissariat a lEnergie Atomique et aux Energies Alternatives CEA
Current assignee: Commissariat a lEnergie Atomique et aux Energies Alternatives CEA
Priority date: 2019-05-28
Filing date: 2020-05-28
Publication date: 2022-04-06
Also published as: FR3096782B1; CN114008489A; KR20220014329A; WO2020239884A1; FR3096782A1; JP2022533815A; US20220252744A1

Abstract

Disclosed is a method for identifying emitting species (S₁ – S_N) emitting X- or gamma radiation in a scene, in which a spectrum of the radiation is provided as input of a first set (CBNN_ID) of a plurality of convolutional neural networks, each convolutional neural network of the first set being associated with at least one atomic species to be identified and having at least one output (CS) indicative of the presence or absence of the atomic species in the scene. Advantageously, a second set (CBNN_PRO) of a plurality of convolutional neural networks makes it possible to determine a signal proportion of each emitting species present in the X- or gamma radiation emanating from the scene. A device for implementing such a method is also disclosed.

Description

Title of the invention: Method and device for identifying atomic species emitting X-rays or

gamma

[0001] The invention relates to a method and a device for identifying species emitting X or gamma radiation, and preferably for quantitatively determining the contribution of each species to the radiation. It falls within the technical field of nuclear instrumentation, and more precisely of

X and gamma spectroscopy.

[0002] The invention is applicable whenever it is necessary to identify radionuclides and / or atomic species exhibiting X-ray fluorescence in a sample or in an environment: radiochemistry, chemical and radiochemical analyzes, decontamination and dismantling nuclear sites, etc.

[0003] The term “X or gamma radiation” is understood to mean electromagnetic radiation of energy greater than 100 eV; more particularly, in the context of the invention, we are interested in radiation of energy between about 2 keV and 2 MeV.

The distinction between X-ray and gamma radiation is not due to the nature of the

radiation, but at its origin: an X-ray radiation has an origin

electronic (typically electronic transitions involving internal energy levels) while gamma radiation is nuclear in origin. Also, the invention makes it possible to identify radionuclides (isotopes) from their gamma radiation spectrum and atomic species that are not necessarily radioactive from their X-ray fluorescence spectrum. The atomic species can be identified from their spectrum. X-ray fluorescence and

radionuclides that can be identified from their gamma spectrum are referred to together as "emitting species". In the following, we will consider more specifically the case where the emitting species are radionuclides but, unless otherwise indicated, all that will be said can also apply to X emitters.

[0004] The conventional methods of identifying radionuclides from their gamma radiation are mainly based on the extraction of the peaks photoelectric in the spectrum resulting from the acquisition of the detector or on the study of areas of interest in the spectrum. See for example (Lutter 2018).

In this type of method, spectral signatures characteristic of

radionuclides of interest (gamma or X lines) are fitted with a Gaussian model and a continuous background to deduce their position in energy. One or more peaks present in the spectrum are compared simultaneously with the tables of nuclear lines of the radioactive isotopes (or with the X-ray fluorescence lines of the various elements), which allows their identification.

These methods have the disadvantage of requiring the use of an expert to identify areas of interest in the spectrum. In addition, they require sufficient photon statistics to clearly show the peaks to be identified. The information contained in the continuous spectrum background (Compton background) is lost.

[0005] Such methods can be accompanied by artificial intelligence techniques, and in particular neural networks. Already in 1995 (Vigneron 1995) had proposed to analyze gamma emission peaks by means of a neural network of the "multilayer perceptron" type to determine the rate

enrichment of uranium. More recently, (Yoshida 2002) has applied a "multilayer perceptron" neural network to the identification of radionuclides in a mixture, and (Medhat 2012) to quantitative radiochemical analysis. However, the use of artificial intelligence techniques does not make it possible to overcome all the drawbacks of these approaches, and in particular the need for "expert" preprocessing to determine the regions of interest of the spectra to be analyzed.

[0006] Other techniques make it possible to analyze an entire spectrum without the need for expert preprocessing.

[0007] For example (Olmos 1991) proposes to use a neural network to identify radionuclides in a mixture from the entire spectrum of gamma radiation emitted by the latter. The article does not specify the type of neural network actually used; in addition, the required signal level must be relatively high (approximately 10 ⁴ photons detected in the main emission peak alone). [0008] (Bobin 2016) uses impulse neural networks (“spiking neural networks”) to determine the proportion of radionuclides in a mixture from the spectrum of gamma radiation emitted by the latter. The method requires a precise model of the source spectrum and is not robust to changes in configuration (variations in the rate of attenuation / scattering of the radiation). In addition, it does not make it possible to identify the radionuclides: also, it is not possible to know whether the atomic species found in low proportion are absent or are really present in low proportions.

[0009] (Kamuda 2017) uses a perceptron-type neural network, trained from synthetic data, to determine the part of the contribution of each radionuclide to the gamma radiation emitted by a source. Intrinsically, the method only makes it possible to identify a limited number (for example 4) of

distinct radionuclides.

[0010] (Abdel-Aal 1997) uses abductive AIM-type neural networks to determine the relative intensities of several sources from low-resolution spectra, in which regions of interest are identified so

automatic. As in (Bobin 2016), we cannot know if the atomic species found in low proportion are absent or are really present in low proportion.

[0011] US 2019/034786 discloses, in general and with few details, the use of a multilayer perceptron to detect or identify

radionuclides. The perceptron can present several outputs, each corresponding to an emitting species or to a group of emitting species.

[0012] CN 109,063,741 also discloses the use of neural networks to detect or identify radionuclides. More specifically, the document discloses the conversion of spectra into two-dimensional images by means of a Hilbert curve before the application of a neural network.

[0013] The invention aims to overcome the aforementioned drawbacks of the prior art. More particularly, it aims to allow the identification of any number of emitting atomic species in a mixture from the X or gamma spectrum of the latter, and this regardless of the configuration of the scene (presence of absorbing or diffusing material between the source (s) present in environment and spectrometric detector) and without sophisticated pre-processing including the identification of regions of interest in the spectrum.

[0014] Advantageously, in addition, the invention aims to determine the probability of the presence of each of the emitting species, and preferably an uncertainty about this probability.

Advantageously, moreover, the invention aims to determine the proportion in the signal of each identified source, as well as to provide an uncertainty on this proportion.

Advantageously, in addition, the invention aims to allow the use of different types of spectrometric detector (CdTe, CdZnTe, Hgl ₂ , Nal, HPGe or any type of gamma spectrometer operating in the energy band from keV to

MeV ...), provided that it is capable of measuring and restoring the energy of each event detected, event by event, or at least, of providing the spectrum of the energies measured.

[0017] Advantageously, moreover, the invention aims to make it possible to avoid long and complex fine-energy calibration operations and to overcome possible drifts of the detector or of operational conditions over time. Ideally, a factory calibration with an error of around 2% should be sufficient to identify the emitting species and, if necessary, determine their proportion.

[0018] In accordance with one aspect of the invention, at least some of these goals are achieved through the use of a plurality of convolutional-type neural networks, each responsible for identifying a single emitter species. Alternatively, as will be discussed in more detail below, each neuron could be tasked with identifying a distinct group of emitter species.

According to another aspect of the invention, at least some of these goals are achieved through the use of a second plurality of convolutional type neural networks, each responsible for determining the proportion of a only emitting species (or group of emitting species) already identified.

[0020] According to another aspect of the invention, these neural networks are trained from synthetic spectral data. In accordance with another aspect of the invention, the spectra supplied as input to the neural networks for identifying the emitting species are previously converted to the logarithmic scale.

According to another aspect of the invention, a "dropout" operation (random extinction of a fraction of the neurons of the inner layers) is applied to these neural networks in order to determine the levels of uncertainty on the presence. , and where appropriate on the proportion, of each emitting species.

According to another aspect of the invention, the different neural networks each have a pair of complementary output neurons, for example using an activation function of the "softmax" type.

[0024] According to another aspect of the invention, once the emitter species have been identified, this information is used to perform an energy "self-calibration" of the detector.

[0025] An object of the invention is therefore a method for identifying species emitting X or gamma radiation in a scene, the method comprising the following steps:

a) acquire, by means of a spectrometric detector, a spectrum of X or gamma radiation from the scene;

b) applying to the acquired spectrum a first data transformation operation including at least one normalization;

c) providing the transformed spectrum as the input of a first set of a plurality of convolutional neural networks, each convolutional neural network of said first set being associated with a respective emitting species to be identified, or with a group of emitting species to identify respective and having at least one output; and

d) for each convolutional neural network of the first set, determine whether the corresponding emitter species, or the group of emitter species

corresponding, is present in the scene as a function of said one or more outputs, steps a) to d) being implemented by means of a signal processing circuit.

[0026] Another object of the invention is a computer program product

comprising instructions which, when the program is executed by a computer, lead the latter to implement steps b) and following of such a method.

[0027] Yet another object of the invention is a device for identifying species emitting X or gamma radiation in a scene, comprising:

a circuit for processing signals generated by a spectrometric detector, said circuit being configured or programmed for:

acquiring from said detector a series of events, each event being associated with a physical quantity representative of an energy value of an X or gamma photon detected by said spectrometric detector;

converting said series of events into an X or gamma radiation energy spectrum by applying a calibration function dependent on a set of calibration parameters;

applying to the energy spectrum of X or gamma radiation a first data transformation operation including at least one normalization;

providing the spectrum thus transformed as the input of a first set of a plurality of convolutional neural networks, each convolutional neural network of said first set being associated with a respective emitting species, or with a respective group of emitting species, and having at least one outlet; and

for each convolutional neural network in the first set, determine whether the corresponding emitter species or the group of emitter species

corresponding is present in the scene as a function of said output or outputs.

[0028] According to particular embodiments of such a device:

The signal processing circuit generated by the radiation detector can also be configured or programmed for:

applying to the energy spectrum of X or gamma radiation a second data transformation operation including at least one normalization;

providing the spectrum thus transformed as the input of a second set of a plurality of convolutional neural networks, each convolutional neural network of said second set being associated with one or more respective emitting species having been determined as being present in the scene and having at least one outlet; and

for each convolutional neural network of the second set, determine as a function of said scalar output or pair of scalar outputs, a signal proportion of the species or of the corresponding emitting species.

The signal processing circuit generated by the radiation detector can also be configured or programmed to determine optimal values of said calibration parameters by maximizing a correlation function between an acquired spectrum and a theoretical spectrum calculated as a function emitter species determined to be present in the scene.

Each convolutional neural network can be associated with a single respective emitter species.

Each convolutional neural network can comprise a pair of complementary output neurons.

The X or gamma photons detected can exhibit an energy in at least part of the range between 2 keV and 2 MeV.

The device can also include said spectrometric detector (SPM).

[0035] The accompanying drawings illustrate the invention:

[0036] [Fig. 1] is a block diagram of a device according to one embodiment of the invention.

[0037] [Fig. 2] is a representation of a convolutional neural network that can be used in a device and / or a method according to the invention.

[0038] [Fig. 3] is a flowchart of a method according to one embodiment of the invention.

[0039] [Fig. 4] is an example of an X or gamma radiation spectrum.

[0040] [Fig. 5] illustrates the identification of atomic species at the origin of

radiation of [Fig. 4]

[0041] [Fig. 6] illustrates the determination of the proportions of these atomic species.

[0042] [Fig. 7 A],

[0043] [Fig. 7B],

[0044] [Fig. 7C], [0045] [Fig. 7D],

[0046] [Fig. 7E] and

[0047] [Fig. 7F] illustrate the benefit of using neural networks with two complementary outputs rather than a single output.

[0048] [Fig. 8A],

[0049] [Fig. 8B], and

[0050] [Fig. 8C], illustrate the benefit of using separate neural networks for the identification of atomic species and for the quantification of their contribution to radiation.

[0051] The device of [Fig. 1] comprises a spectrometric detector of

X or gamma radiation (i.e. a detector sensitive to the energy of the photons detected) and a CTS processing circuit for signals generated by the radiation detector.

The SPM spectrometric detector comprises a sensitive element ES, preferably pixelated, an analog reading circuit EL and a converter

analog-to-digital ADC.

The spectrometric detector acquires photons from an SC scene in which are located different radionuclides (or atomic species emitting X fluorescence radiation) S, (Si ... SN) of activity A, (Ai ... AN) - whose identity and relative abundance are unknown a priori - potentially located at different distances from the detector. Absorbent or diffusing ABS material may lie between one or more sources and the detector. Each radionuclide S, emits photons P, at energies E _k , i. For example, Pi (E _{k, i} ) denotes a photon of energy E _{k, i} emitted by the first radionuclide Si.

The sensitive element ES can be of any type suitable for detecting X / gamma photons from the scene in at least a portion of the range

spectral 2 keV - 2 MeV. It may be, for example, a semiconductor pixel made of Si, Ge, CdTe, etc., a scintillator sensor, a perovskite sensor, etc.

The sensitive element generates a signal in the form of a physical quantity, generally electrical, representative of the energy of each X or gamma photon received (typically, it is a current pulse whose charge electric is proportional to this energy). The EL reading electronic circuit performs conventional analog preprocessing of the signals coming from the sensitive element: amplification, pulse shaping, detection of their height or energy. The analog signals SA from the readout electronics are converted into digital format by the ADC converter.

Preferably, the photons are detected one by one, their energy is recorded and dated by the detector. This information, contained in the digital data stream FDN coming from the converter ADC, is transmitted to the processing circuit CTS. The latter can be on board or deported; in the latter case, a telecommunications link must be established between the spectrometric detector and the processing circuit.

The signal processing circuit can comprise one or more generic processors, or ones specialized for the digital processing of signals, programmed in a timely manner. As a variant or in addition, it can include dedicated digital circuits. In general, moreover, it includes random access memories to store the data to be processed (in particular the events generated by the spectrometric detector) and active and / or dead memories to store calibration parameters, neural network coefficients, etc. . In general, the invention is not limited to a particular technology for making the signal processing circuit. In the following description of this circuit, the breakdown into blocks and modules is purely functional, these blocks and modules do not necessarily correspond to distinct physical elements.

[0058] In the embodiment of [Fig. 1], the signal processing circuit CTS comprises three modules: a source identification module ID, a learning module APP and a self-calibration module AE. In other embodiments, the training module may be absent, in which case the training of the neural networks of the identification module is performed by means of another device and the coefficients of the learned neural networks are simply transferred to the issuing species identification device. In other embodiments, the self-calibration module may be absent, but this means that an accurate calibration must be performed beforehand (for example, in a laboratory metrology) and that precautions should be taken to minimize drifts in the response of the spectrometric detector.

The FDN digital data stream from the spectrometric detector is received by the identification module of the CTS circuit and stored in an MEV event memory. An MCS spectra building module converts these events into spectra using a TE calibration table, stored in memory, which allows photon energy to be associated with each detection event. This calibration table, established beforehand, can be relatively imprecise, with an error in the energy values which can reach 2%. As will be explained in detail below, the AE auto-calibration module allows updating of calibration tables improving their accuracy. In the case of a pixelated detector, the calibration is done pixel by pixel, with a different calibration table for each pixel.

[0060] Each spectrum is in fact an energy histogram: an energy value is assigned to each event; the events are grouped into energy classes ("bins") and the spectrum is made up of the number of events belonging to each class. The term "spectrum" therefore means the spectral distribution of the photons detected over a given time interval, the duration of which can be fixed or chosen by the user.

A module for transforming MTD data then performs preprocessing of the spectra. In the embodiment considered here, two separate pre-treatments are carried out.

On the one hand, each acquired spectrum is converted to the logarithmic scale, then normalized. The spectrum thus transformed, SNlog, is used for the identification of the emitting species. More specifically, let Si be the number of events in the i-th energy class. We apply a first operation

[0063] [Math. 1]

[° ⁰⁶⁴ ] ^SN Î = ' ^O “fe)

Then the normalized spectrum is calculated as follows:

[0066] [Math. 2]

[0068] The use of a logarithmic scale makes it possible to reveal spectral structures of low amplitude but which effectively contribute to the identification of emitting species.

On the other hand, the spectrum is also normalized to standard 1 (which consists in ensuring that the sum of the values associated with each energy class is 1) without logarithmic conversion. The spectrum thus transformed, SN1, given by:

[0070] [Math. 3]

[0072] is used to determine the proportions of the identified emitting species.

[0073] More generally, the preprocessing can be a transformation of scale. In any case, and contrary to the teaching of CN 109 063 741, there is no change in the dimensionality of the data - the preprocessed spectra remain one-dimensional.

The SNlog spectrum is supplied as input to a CBNNJD module which implements a plurality of M “bayesian” convolutional neural networks, each taking the entire spectrum as input and providing at its output a value PPj indicative of a probability of the presence in the scene of a particular radionuclide (identified by the index "j"), as well as a level of confidence in this probability. There is therefore a separate neural network for each emitter species that we want to be able to identify. The structure and operation of these Bayesian convolutional neural networks will be described in more detail below.

An MS thresholding module is used to determine which emitting species are considered to be actually present in the scene. For this, the thresholding module takes into account the probability of presence and, possibly, its level of uncertainty.

At the output of the identification module, we therefore obtain an LEA list of emitting species present in the scene. The spectrum normalized in standard 1, SN1, is supplied as input to a CBNN_PRO module which implements a plurality of M “bayesian” convolutional neural networks, each corresponding to a particular radionuclide. Each neural network of the CBNN_PRO module which corresponds to a radionuclide “j” already identified as being present in the scene takes as input the entire spectrum and provides at its output a value PROj indicative of the proportion of this radionuclide in the recorded signal , as well as a level of confidence in this proportion.

In other words, the value PROj corresponds to the percentage of recorded photons which can be attributed to the radionuclide "j" (by abuse of language one speaks more simply of the proportion of the radionuclide "j" in the scene, but this is only correct in special conditions, if all the emitters are at the same distance from the detector and in the presence of the same absorbent / scattering materials).

So there is a separate neural network for each atomic species. These neural networks can be of the same type as those used for

the identification of emitter species.

At the output of the identification module, we therefore also obtain an LP list of proportions of the identified atomic species.

The neural networks of the CBNNJD and CBNN_PRO modules were previously trained by supervised learning from a synthetic BDS database, that is to say simulated data, generated by the APP learning module. The learning produces two parameter databases, PRNJD and PRN_PRO, characterizing neural networks

identification and determination of proportions, which are stored in memories.

The synthetic spectra are created by a Monte-Carlo simulator (SS block in [Fig. 1]) which makes it possible to simulate the photon-matter interactions in the detector as well as in the direct environment of the detector and, advantageously, possible background noise. For each photon, it is then necessary to apply the response of the detector, that is to say, the energy resolution due to statistical fluctuations in the creation of electron-hole pairs, that due to electronic noise as well as the loss. load in the detector (RD block). It is a physical model which is applied only once for a list of given source of interest comprising as many sources as desired. Each emitting atomic species is simulated independently and the simulated data are restored in the form of a list of events giving the energy deposited by each photon in the spectrometric detector. Then, mixtures of different radioelements are also generated synthetically (MIX block): for this, energies are randomly drawn from the energy lists simulated for each radioelement in the mixture, with different statistics and different proportions. The proportion of photons attributed to each emitting atomic species is recorded.

[0082] An intentional "de-calibration" is then applied so that the neural networks learn the effect of a drift of the calibration laws. We derive a gain g in a Gaussian centered on 1 and standard deviation ogain as well as an offset off in a Gaussian centered on 0 and standard deviation ooff. On all the energies E of the same mixture, we then calculate the new de-calibrated energy Ed by:

[0083] [Math. 3]

[0084] E = g (E— off)

Several de-calibrated synthetic spectra of each source and source mixture are recorded in the BDS database used for

learning.

The AJD1 - AJDM blocks implement the supervised learning algorithms of the M convolutional neural networks of the CBNNJD block, and produce M sets of parameters characterizing each of these neural networks; these parameters are stored in the aforementioned BDNJD database. Likewise, the A_PR01 - A_PROM blocks implement the supervised learning algorithms of the M convolutional neural networks of the CBNN_PRO block, and produce M sets of parameters characterizing each of these neural networks; these parameters are stored in the aforementioned BDN_PRO database.

The synthetic sources of the SS block are also used by the AE module in order to implement a self-calibration process of the spectrometric detector. By knowing the radionuclides present in the source and their proportions (data provided by the ID identification module), it is indeed possible to use these synthetic sources to calculate an “expected” spectrum. A An adaptive mesh or genetic algorithm type algorithm is then used by the ECC block to find a set of calibration parameters which maximizes the correlation between this expected spectrum and that supplied by the MCS module. These parameters are used to update the calibration tables used by the ID module. Note that some blocks (MEV, MCS, SS, TE) appear in several places in [Fig. 1] for the sake of readability.

The ECC block implements the energy calibration algorithm by correlation described in (Maier 2016).

This calibration does not require user intervention and is based only on the measurement carried out in real time in the scene to be analyzed

The Bayesian convolutional networks used for the identification of emitting atomic sources and the determination of their proportions will now be described with the aid of [Fig. 2]

In a manner known per se - see for example (Aloysius 2017) - a convolutional network comprises a first part composed of several convolution layers making it possible to extract different characteristics of the input spectra and a second part consisting of a multilayer perceptron (in the English language literature we also speak of “fully-connected” layers, that is to say

“Fully connected”).

In the embodiment of [Fig. 1], the input is an SP0 spectrum comprising 2000 channels, each channel being a representative value of the spectrum energy in a respective spectral range.

The first convolutional layer CC1 comprises 16 convolutional neurons. Each convolutional neuron does the following:

- filtering by a convolutional filter having a kernel of 16 elements, with a padding by zeros to preserve the dimension of the data.

- Dimensional reduction of the filter output by an operation of "Max

Pooling ”size two.

- Batch normalization.

- Application of a non-linear activation function, in this case of the ReLU type. The size 2 MaxPooling operation consists of taking the spectrum and keeping only one channel out of two, the largest. This reduces the dimension of the spectrum by a factor of 2.

The batch normalization operation - known per se, see (loffe 2015) - consists in gathering the data from the learning database by subsets called batches or "batches", performing an iteration of learning on each of the batches (as will be described later), then normalize on average and in variance the outputs of each neuron corresponding to the batch considered. After training, the normalization settings across the database are saved to apply the same normalization when using neural networks for identifying and determining the proportions of emitter species.

The ReLU activation function is defined by:

ReLU (x) = max (0, x).

The output of the first convolutional layer therefore consists of 16 spectra of characteristics SP1, 0 - SP1, 15 of dimension 1000. These data are supplied as input to the second convolutional layer CC2 which is in all respects similar to the first , except in that it operates on data of dimension 1000.

The output of the second convolutional layer therefore consists of 16 spectra of characteristics SP2.0 - SP2.15 of dimension 1000. These data are supplied as input to the third convolutional layer CC3 which is in all respects similar to the first two. , except in that it operates on data of dimension 500 and that its output therefore consists of 16 spectra of

characteristics SP3.0 - SP3.15 of dimension 250.

The latter are flattened to form a SPA4 vector of 4000 elements which are all supplied as input to each of the 20 neurons of a perceptron layer CP using, like the convolutional layers, a ReLU type activation function.

The last layer, or output layer, CS of the perceptron is

preferably composed of two neurons with an activation function of the “softmax” type where, by noting X _j the output of the neuron before the activation function: [0101] [Math. 4]

[01 02] softmax (

[0103] j being the index which identifies the two output neurons.

[0104] For identification networks, the first neuron of the output layer gives a number between 0 and 1 which represents the absence or presence of the radioelement in the mixture, while the second neuron is the

complementary to 1 of the first neuron.

[0105] For proportion evaluation networks, the first neuron gives a number between 0 and 1 which corresponds to the signal proportion of the

radioelement while the second neuron is the signal proportion of all other elements.

[0106] The use of two complementary output neurons is not essential, but advantageous as will be discussed later with the aid of Figures 7A to 7F.

The parameters or coefficients of the neural networks (cores of the convolutional layers CC1, CC2 and CC3, synaptic weights of the perceptron layers) are learned, in a supervised manner, using for example a gradient descent algorithm. For example, in the embodiment discussed later with reference to Figures 4 to 6, a stochastic gradient descent algorithm is used with a learning rate of 0.01 which decreases by 0.001 at each iteration, a moment of 0.9 and the use of the Nesterov moment, the training being carried out over 10 iterations on batches of 1000 examples.

[0108] More specifically, in this embodiment, the cost function used and minimized during the learning phase for this identification network is binary cross entropy, defined as:

[0109] [Math. 6]

[01 10] _(pred y, y _r ERE) ^- yréel · lo§ (ypred) ^"F (1 ^- yréel) · log (l ^- Real y)

[01 11] where yreal is the true value that the neuron must have at the output and ypred is the value predicted by the network.

[01 12] The learning of the ratio evaluation network is carried out with the synthetic spectra normalized to standard 1. The output to be predicted by the first neuron is the proportion of the radioelement in the mixture and that of the second neuron is the proportion of the other radioelements.

[01 13] The cost function used and minimized for this network is the root mean square deviation:

[01 14] [Math. 5]

2

[01 1 5] L (y _pred , y _réei ) - (y _pred ^- y real)

[01 16] The two cost functions are evaluated on average over all the examples in the database.

[01 17] The “Bayesian” character of the neural network of [Fig. 2] is obtained by randomly switching off a fraction of the neurons in the layers

intermediaries (“dropout”), see (Gai 2016), and this as well during

learning only while using the network. Each neural network is applied a plurality of times to each input spectrum and, due to the random extinction of the neurons, each time it returns a different result, the statistical distribution of which provides information on the uncertainty affecting the identification of d 'an emitting atomic species and / or its proportion. For example, in the embodiment discussed later with reference to Figures 4 to 6, the rate of neuron extinction is 50% for each of the intermediate layers and each neural network is applied 100 times.

[01 18] [Fig. 3] schematically illustrates the process for identifying emitting atomic species described above.

[01 19] Step a) comprises acquiring a spectrum using the SPM spectrometric detector.

[0120] Step b) comprises the transformation of the data prior to the application of the identification neural networks - that is to say, in a preferred embodiment of the invention - the conversion of the spectrum to the logarithmic scale and its normalization.

[0121] Step c) comprises the application of convolutional neural networks, preferably Bayesian, to the spectrum transformed by step c). [0122] Step d) comprises the identification of the emitting species present in the scene as a function of the outputs of the identification neural networks.

Step e) comprises the transformation of the data prior to the application of the proportion neural networks - that is to say, in a preferred embodiment of the invention - the normalization of the spectrum into a standard 1.

[0124] Step f) comprises the application of convolutional neural networks, preferably Bayesian, to the spectrum transformed by step e).

[0125] Step g) comprises determining the proportions of the emitting species identified in step d) according to the outputs of the proportion neural networks.

[0126] Step h) corresponds to the self-calibration, which updates the tables

calibration methods used for carrying out step a).

[0127] The preliminary step app) corresponds to the learning of the neural networks used in steps c) and f).

[0128] The process can stop at step c) if only the identification of the emitting species is required, or even at step g) if self-calibration is not used.

[0129] The invention has been tested using, as a spectrometric detector, a pixelated spectro-imager based on CdTe semiconductor and its optimized read circuits (Gevin 2012). The spectro-imager comprises 256 pixels at 800 µm pitch with 2 mm thickness, an energy dynamic of 1 keV to 1 MeV, and an energy resolution of 0.7 keV (full width at half height) at 60 keV.

Similar results were obtained using a spectro-imager of the same type but having a thickness of 1 mm and a pitch of 625 μm. The change of detector did not even require relearning the neural networks, which confirms the robustness of the identification method according to the invention.

The spectro-imager was installed in a vacuum chamber, cooled by Peltier modules to a temperature of approximately -15 ° C. and subjected to a polarization field of 300 V / mm. The signal recorded by the read circuits is transmitted to a Zynq programmable digital circuit via a CIF interface card. The digital circuit stores the events and transmits them to a control computer on which the spectro-identification process is carried out. The events are returned as a list containing the interaction pixel, the uncalibrated deposited energy and the interaction date. The sum spectrum of all the pixels is pre-calibrated with a calibration table determined in advance by a unique initial calibration in the laboratory (as a variant, the theoretical transfer function of the acquisition chain could be applied).

[0132] The synthetic data were generated by means of semi-analytical simulation by modeling the detector and its very close environment, namely the structure which contains the detector. Photon-matter interactions in

the environment and in the detector were simulated and the energy deposits in the detector and their position were recorded. A Geant4 simulator can also be used to perform these simulations.

[0133] For each energy deposit, the statistical energy fluctuations were taken into account by deriving an energy in a Gaussian centered on the energy E0 deposited, of standard deviation

[0134] [Math. 7]

[0135] s = / s _w E ₀ F, WHERE

[0136] ew = 4.42 eV is the energy for creating electron-hole pairs in CdTe and F = 0.15 is the Fano factor for CdTe.

[0137] To take into account the pressure drop, the weighting potential in the detector was calculated by solving the Poisson equation AUw = 0 in cylindrical coordinates on the detector using the finite difference resolution method , with the boundary condition of setting 1 on the electrode on which the signal is induced and 0 on the other electrodes and on the cathode. The simulation domain has been extended to 2 electrodes on either side of the electrode on which the signal is induced. As a variant, a finite element resolution method could have been used. Then, the CCE pressure drop was calculated as a function of the depth zO using the Hecht equation:

[0138] [Math. 8]

[0139] CCE (z ₀ ) = f ^L Ekp ftp

[0140] where, pexe = 1, 14.10-3cm ² .V-1, phxh = 3.36.10-4cm ² .V-1, L = 2 mm and V0 = 600 V. [0141] The ECCE energy at the interaction depth zO was calculated by

[0142] [Math. 9]

CCE (z ₀ )

[0143] E CCE = E _f .

max _z (CCE (z)) ^'

[0144] It would also have been possible to calculate the pressure drop in 3D in the detector and not only in 1 D to obtain a more faithful model of the response of the detector which is however not essential for learning the network. neurons.

[0145] The error linked to electronic noise was taken into account by drawing the final energy recorded in a Gaussian centered on ECCE with a standard deviation s = enn rms where rms = 50 e-corresponds to the number of electrons rms due to electronic noise.

[0146] To create the lists of mixtures of atomic species, between 10 and 10 million photons were drawn from each event list of each individually simulated radioelement, drawing carried out with replacement, which means that the same energy can be drawn several times. The de-calibration was performed by taking ogain = 0.0055 and ooffset = 0.5 keV. In this way, 200,000 synthetic blends were made.

The synthetic spectra were obtained by making a histogram composed of 2000 channels of 0.5 keV in width on an energy dynamic of 0 to 1 MeV.

[0148] The neural networks used for the identification and determination of the proportions are as described above with reference to [Fig. 2]

[0149] [Fig. 4] shows a spectrum (shown on a logarithmic scale) obtained by exposing the spectro-imager to a mixture comprising 57Co and 137Cs. Note that the number of photons detected is relatively low, in the order of a few thousand.

[0150] [Fig. 5] illustrates the output values of several identification neural networks, corresponding to these nuclides but also to others which are not present in the scene (241 Am, 133Ba, 152Eu, 22Na). Specifically, as each neural network is applied 100 times by randomly switching off 50% of the neurons at each repetition, we obtain a statistical distribution of the outputs. The vertical bars represent the medians, the error boxes correspond to the first and third quartiles, and the error bars to the first and ninth deciles.

[0151] The detection threshold is set at 0.5 on the median, which is natural if we consider that the output of an identification network represents the probability of the presence of the corresponding emitting species. We note that the statistical dispersion of the outputs is negligible for the radionuclides actually present and, for the others, sufficiently low not to create any risk of false positive.

[0152] [Fig. 6] illustrates the statistical distribution (median and error bar corresponding to the first and ninth deciles) of the proportions of 57Co and 137Cs (more precisely, of their contributions to the total number of photons detected), obtained by applying 100 times the networks of proportion neurons

corresponding, by switching off 50% of the neurons each time.

[0153] Figures 7A to 7F illustrate the advantage provided by the use of neural networks having two complementary outputs with an activation function of the softmax type. In these figures, the continuous line curves represent the rate of correct responses (1: absence of identification errors) in the case of a softmax activation function as a function of the number of photons detected, while the curve in hatched line corresponds to single output neurons, with a sigmoid activation function (the softmax function cannot be used with a single output neuron, it would always return 1). Each point in these graphs corresponds to the average rate of correct answers over 2,500 mixtures of sources. The different graphs correspond to different radionuclides: 241 Am ([Fig. 7A]), 133Ba ([Fig. 7B]), 22Na ([Fig. 3C]), 152Eu ([Fig. 7D]); 137Cs ([Fig. 7E]), 57Co ([Fig. 7F]).

The results show that the performances of networks with 1 output neuron are equivalent to those of networks with 2 output neurons in the case of sources of 241 Am and 152Eu. In the case of 57Co and 22Na sources, the relative performance depends on the number of photons, and in the case of 22Na, the performance of the 1-neuron output network very little exceeds that of the network with 2 output neurons. Finally, in the cases of 133Ba and 137Cs, the 2-output neural networks outperform the 1-output neural network.

[0155] In general, at least in the configuration tested, it appears preferable to use a neural network with 2 outputs.

In all cases, it is noted that the rate of correct responses is close to or greater than 0.9 as soon as the number of photons reaches 103.

[0157] An important aspect of the invention is the separation of the steps of identifying the emitter species and determining their proportions, which are carried out by distinct sets of neural networks.

[0158] This makes it possible to use different pretreatments of the spectra for the two operations. For the identification, a "logarithmic normalization" is preferably used which makes it possible to give importance even to structures which are composed of few photons, which is in particular interesting because the probability of interaction of the photons in the detector decreases sharply when the energy of the photons increases. For the proportion evaluation network, on the other hand, the invention preferably uses linear normalization (in norm 1) which "ignores" small structures. This is advantageous because the proportions of each source in the signal are linearly related to their contribution in the spectrum. Experimentally, it has been found that using a "log normalization" for the evaluation of proportions does not give good results.

[0159] This is illustrated by the figures [Fig. 8A], [Fig. 8B] and [Fig. 8C].

[0160] The spectrum of [FIG. 8A] was analyzed by the method of the invention. The [Fig. 8B] shows that the identification networks can confirm with certainty the presence of 241 Am and 57Co. Assessment networks

However, proportions give a proportion of 0.7% for 57Co and 99.3% for 241 Am.

If only the proportions had been determined, as in certain methods of the prior art, it would have been impossible to know if 57Co was really present in the trace state, or it was in fact absent, such as, for example, 133Ba, 137Cs and 22Na. The use of identification networks distinct from the proportion makes it possible to unambiguously determine that 57Co is actually present, albeit in small proportion.

[0162] The method and device of the invention have several advantages over the prior art.

[0163] The use of a separate neural network for each source that we want to identify makes it possible not to restrict the number of atomic species that we are able to identify. Adding new sources, moreover, is possible by adding new networks without changing the overall architecture and by relearning only for the perceptron part. It is not necessary, however, to re-train on the convolutional layers.

[0164] As a variant, it is possible to associate a neural network with a plurality of emitting species, for example a family of species exhibiting similar spectra. In this case, it is not individual species that are identified, but groups or families of species.

The use of convolutional-type neural networks confers great robustness with respect to calibration errors (temporal instability of the calibration law, variable instrumental responses from one spectrometric detector to another) as well as configuration changes (presence of absorbent or diffusing materials between the source (s) and the detector (s)

spectrometry) by taking these issues into account in the database. In addition, it makes it possible to process cases of spectra with low photon statistics with good performance (accuracy greater than 90% with spectra containing at least 1000 photons resulting from the mixture of sources). This method does not focus on particular peaks but on the general structure of the spectra (peaks, Compton structures in particular).

[0166] The Bayesian approach makes it possible to quantify an error in the probability of the presence of each source, which informs the user of the relevance of the results and therefore of the degree of confidence that he can grant to an automatic analysis.

[0167] In addition, no “expert” pre-processing is necessary, and the fine calibration is carried out automatically. The use of qualified technicians is therefore greatly reduced. As explained above, the separation of the steps for identifying and determining the proportions makes it possible to reliably determine whether an atomic species is absent or indeed present but in a small proportion.

[0169] The invention has been described with reference to a particular embodiment, but many variants are possible.

[0170] For example, concerning the structure of neural networks, the number of convolutional and perceptron-type layers, the activation functions, the data reduction method, the size of the convolutional nuclei, etc. are given as an example only. Moreover, all neural networks do not have to be the same.

[0171] Other supervised learning techniques can be implemented. In addition, the use of synthetic sources for learning is advantageous, but it is also possible to use - in replacement or in addition to real sources.

[0172] Other data transformation operations prior to the application of neural networks, and in particular other normalization techniques, can be used.

[0173] Recourse to a random extinction of neurons is not essential if one does not wish to have information on the uncertainty of the measurements. In addition, when used, the random extinction does not necessarily have to affect all of the intermediate layers.

[0174] Criteria other than a threshold of 0.5 can be used to determine whether an atomic species is considered to be present or absent, in particular if it is preferred to minimize the risk of false positives or, conversely, of false negatives.

Each neural network can have more than two outputs - for example an output indicative of the probability of the presence of an emitting species, an output indicative of the probability of its absence and an output indicative of an indeterminate situation. Advantageously, these outputs are complementary (that is to say that their sum takes a fixed value, typically 1). The activation function outputs can be of the Softmax type, but other possibilities can be envisaged by those skilled in the art.

[0176] Data originating from several distinct spectrometric detectors can be combined.

[0177] Bibliographical references

[0178] (Lutter 2018) G. Lutter, M. Huit, G. Marissens, H. Stroh, F. Tzika "A gamma-ray spectrometry analysis software environment" Applied Radiation and Isotopes 134 (2018): 200-204.

[0179] (Vigneron 1995) V. Vigneron, J. Morel, M.-C. Lepy, J.-M. Martinez "Statistical modeling by neural networks in gamma-spectrometry" Conference and

International Symposium on Radionuclide Metrology (1995).

[0180] (Yoshida 2002) E. Yoshida, K. Shizuma, S. Endo, T. Oka "Application of neural networks for the analysis of gamma-ray spectra measured with a Ge

spectrometer »Nuclear Instruments and Methods in Physics Research A 484 (2002).

[0181] (Medhat 2012) M. E. Medhat “Artificial intelligence methods applied for quantitative analysis of natural radioactive sources”, Annals of Nuclear Energy 45 (2012).

[0182] (Olmos 1991) P. Olmos, J. C. Diaz, J. M. Perez, P. Gomez, V. Rodellar, P. Aguayo, A. Bru, G. Garcia-Belmonte, J. L. de Pablos "A New Approach to

Automatic Radiation Spectrum Analysis »IEEE Transactions On Nuclear Science, Vol. 38, No. 4 (1991).

[0183] (Bobin 2016) C. Bobin, O. Bichler, V. Lourenço, C. Thiam, M.Thévenin “Real- time radionuclide identification in g-emitter mixtures based on spiking neural network” Applied Radiation and Isotopes 109 (2016 ) 405-409 (2016).

[0184] (Kamuda 2017): M. Kamuda, J. Stinnett, C. J. Sulliva "Automated Isotope Identification Algorithm Using Artificial Neural Network" IEEE Transactions On Nuclear Science, Vol. 64, No. 7 (2017).

[0185] (Abdel-Aal 1997) R.E. Abdel-Aal, M.N. Al-Haddad "Determination of radioisotopes in gamma-ray spectroscopy using abductive machine learning"

Nuclear Instruments and Methods in Physics Research A 391 (1997). [0186] (Maier 2016) D. Maier, O. Limousin “Energy calibration via correlation”, Nuclear Instruments and Methods in Physics Research, Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, vol. 812, p. 43-49. (2016).

[0187] (Gai 2016) Y. Gai, Z. Ghahramani “Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning”, Proceedings of the 33 rd International Conférence on Machine Learning, New York, NY, USA, 2016. JMLR : W&CP volume 48.

[0188] (Gevin 2012) O. Gevin, et al. "Imaging X-ray detector front-end with high dynamic range: IDeF-X HD." Nuclear Instruments and Methods in Physics Research Section A: Accelerators ", Spectrometers, Detectors and Associated Equipment 695 (2012): 415-419.

[0189] (Aloysius 2017) N. Aloysius, M. Geetha "A Review on Deep Convolutional Neural Network" International Conference on Communication and Signal

Processing, April 2017, India.

[0190] (loffe 2015) S. loffe, C. Szgedy Geetha “Batch Normalization: Accelerating Deep Network Training by Reducing Internai Covariate Shift” Proceedings of the 32nd International Conference on Machine Learning, 2015, France.

Claims

1. Method for identifying species emitting (Si ... SN) of X or gamma radiation in a scene, the method comprising the following steps:

a) acquire, by means of a spectrometric detector (SPM), a spectrum of X or gamma radiation from the scene;

c) provide the transformed spectrum as the input of a first set

(CBNNJD) of a plurality of convolutional neural networks, each convolutional neural network of said first set being associated with a respective emitter to be identified, or with a group of respective emitter to be identified, and having at least one output; and

d) for each convolutional neural network of the first set, determine whether the corresponding emitting species, or the corresponding group of emitting species, is present in the scene as a function of said output (s).

steps a) to d) being implemented by means of a signal processing circuit (CTS).

2. The method of claim 1 wherein:

each convolutional neural network of the first set includes an input layer (CC1), an output layer (CS) and at least one intermediate layer (CC2, CC3, CP), each intermediate layer comprising a plurality of neurons;

step c) is repeated a plurality of times, randomly turning off, each time, a fraction of the neurons of at least one intermediate layer; and

step d) comprises, for each convolutional neural network of the first set, determining the presence of the corresponding emitting species or group of species in the scene as well as a confidence level of said determination from a statistical analysis of the values taken by said output or outputs during the different repetitions of step c).

3. Method according to one of the preceding claims also comprising the following steps, also implemented by means of a signal processing circuit (CTS):

e) applying to the acquired spectrum a second data transformation operation including at least one normalization;

f) provide the transformed spectrum as an input to a second set

(CBNN_PRO) of a plurality of convolutional neural networks, each convolutional neural network of said second set:

being associated with a respective emitter species, or with a group of respective emitter species, having been determined to be present in the scene following step d); and

having at least one outlet; and

g) for each convolutional neural network of the second set, determine as a function of said output (s) a signal proportion of the species or of the corresponding emitting species.

4. The method of claim 3 wherein:

each convolutional neural network of the second set includes an input layer (CC1), an output layer (CS) and at least one intermediate layer (CC2, CC3, CP), each intermediate layer comprising a plurality of neurons;

step f) is repeated a plurality of times, randomly switching off, each time, a fraction of the neurons of at least one intermediate layer; and

step g) comprises, for each convolutional neural network of the second set, determining the signal proportion of the corresponding emitting species or group of species as well as a confidence level of said determination from a statistical analysis of the values taken by said output or outputs during the various repetitions of step f).

5. Method according to one of the preceding claims wherein each

convolutional neural network is associated with a single respective transmitter species.

6. Method according to one of the preceding claims, in which step b) preserves the dimensionality of the acquired spectrum.

7. The method of claim 6 wherein step b) comprises a logarithmic transformation of the acquired spectrum, followed by its normalization.

8. Method according to one of the preceding claims, also comprising a prior step of supervised learning of convolutional neural networks using simulated X or gamma radiation spectra, corresponding to mixtures of known composition of several emitting species.

9. The method of claim 8 wherein each convolutional neural network comprises an input layer, an output layer and at least one intermediate layer, each intermediate layer comprising a plurality of neurons; and said step of supervised learning is performed by randomly turning off a fraction of the neurons of at least one middle layer.

10. Method according to one of the preceding claims, in which step a) of acquiring a spectrum of X or gamma radiation from the scene comprises:

the acquisition of a series of events, each event being associated with a physical quantity representative of an energy value of an X or gamma photon detected by said spectrometric detector; and

the method also comprising a step h) of determining optimum values of said calibration parameters by maximizing a correlation function between said spectrum and a theoretical spectrum calculated as a function of the emitting species determined to be present in the scene.

1 1. Method according to one of the preceding claims wherein each convolutional neural network comprises a pair of complementary output neurons (CS).

12. Method according to one of the preceding claims, in which the acquired spectrum extends, in whole or in part, in a range of between 2 keV and 2 MeV.

13. Computer program product comprising instructions which, when the program is executed by a computer, lead the latter to implement steps b) and following of a method according to one of the preceding claims.

14. Device for identifying species emitting X or gamma radiation in a scene, comprising:

a signal processing circuit (CTS) generated by a spectrometric detector, said circuit being configured or programmed for:

for each convolutional neural network of the first set, determine whether the corresponding emitter species or the corresponding group of emitter species is present in the scene as a function of said output (s).

15. Device according to claim 14 wherein the signal processing circuit generated by the radiation detector is also configured or programmed for:

provide the spectrum thus transformed as an input to a second set of a plurality of convolutional neural networks, each convolutional neural network of said second set being associated with a respective emitter species, or a respective group of emitter species, having been determined to be present in the scene and having at least one exit ; and

for each convolutional neural network of the second set, determine as a function of said scalar output or pair of scalar outputs a signal proportion of the corresponding emitting species or group of species.

16. Device according to one of claims 14 and 15 wherein the circuit for processing signals generated by the radiation detector is also configured or programmed to determine optimal values of said calibration parameters by maximizing a correlation function between an acquired spectrum and a theoretical spectrum calculated as a function of the emitting species determined as being present in the scene.

17. Device according to one of claims 14 to 16 wherein each convolutional neural network is associated with a single respective emitter species.

18. Device according to one of claims 14 to 17 wherein each convolutional neural network comprises a pair of output neurons

complementary (CS).

19. Device according to one of claims 14 to 18 wherein the X or gamma photons detected exhibit energy in at least part of the range between 2 keV and 2 MeV.

20. Device according to one of claims 14 to 19 also comprising said spectrometric detector (SPM). ]