A multi-stage machine learning algorithm for estimating personal dose equivalent using thermoluminescent dosimeter

In the present age, marked by data-driven advancements in various fields, the importance of machine learning (ML) holds a prominent position. The ability of ML algorithms to resolve complex patterns and extract insights from large datasets has solidified its transformative potential in various scientific domains. This paper introduces an innovative application of ML techniques in the domain of radiation dosimetry. Specifically, it shows the applicability of ML in estimating the radiation dose received by occupational workers. This estimation is expressed in terms of personal dose equivalent, and it involves the utilization of thermoluminescence signals emitted by CaSO4:Dy-based personnel monitoring badges. To estimate personal dose equivalent, three-stage algorithm driven by ML models is proposed. This algorithm systematically identifies the photon energy ranges, calculates the average photon energy, and determines personal dose equivalent. By implementing this approach to the conventional three-element dosimeter, the study overcomes existing limitations and enhances accuracy in dose estimation. The algorithm demonstrates 97.8% classification accuracy in discerning photon energy ranges and achieves a coefficient of determination of 0.988 for estimating average photon energy. Importantly, it also reduces the coefficient of variation of relative deviations by up to 6% for estimated personal dose equivalent, compared to existing algorithms. The study improves accuracy and establishes a new methodology for evaluating radiation exposure to occupational workers using conventional thermoluminescent dosimeter badge.


Introduction
Personnel monitoring is an essential radiation protection measure adopted in organizations dealing with the use of ionizing radiation, such as nuclear power plants, medical facilities, and research facilities.Personnel dosimeters, particularly, the thermoluminescence dosimeters (TLDs), are most commonly used to assess the dose received by radiation workers exposed to ionizing radiation from external sources.Personnel dosimeters often measure radiation exposure in terms of operational quantity which is personal dose equivalent (H p (d)) such as H p (10), H p (0.07), and H p (3) [1,2].A realistic and conservative estimation of protection quantities is ensured by these quantities [3].The accuracy of estimated H p (d) becomes crucial as this dose information serves as the foundation for evaluating the protection status of radiation workers, installations and practices including ensuring compliance with regulatory requirements.
In India, about 0.25 million occupational workers are monitored for external exposure due to gamma and beta radiation using CaSO 4 :Dy-based TLD badges [4].CaSO 4 :Dy exhibits a strong energy-dependent response [5,6].To overcome this, TLD badge comprises of a TLD card with of three CaSO 4 :Dy-Teflon discs loaded into a cassette having filters [7,8].However, the use of filters and the disc geometry of the dosimeter results in significant angular dependence.The energy and angular dependence of the response presents a challenge for the accurate estimation of occupational doses.The dose estimation algorithms presently used, either consider the angle-averaged response or the response at normal incidence to fit the response function [9][10][11][12][13].When dealing with lower and intermediate photon energy, both the angular and energy response changes rapidly.These variations cannot be adequately addressed using traditional, average response-based algorithms.Instead, a more flexible approach is needed, and machine learning (ML) algorithms seem to provide the necessary adaptability to account for these variations due to their capability to map complex relationships.
One of the earliest demonstrations of the potential of ML in TL dosimetry was by Moscovitch et al [14], who used an artificial neural network (ANN) to estimate doses in LiF:Mg:Ti-based four-element dosimeters.More recently, researchers have demonstrated the applicability of ML algorithms in TL dating, identification of anomalies in TL glow curves (GC), and classification of thermoluminescence features of natural halite [15][16][17][18][19][20][21][22][23].In the present work, we aimed to develop an algorithm for estimating the average photon energy and the dose in terms of H p (d) from TL readouts of a three-element CaSO 4 :Dy dosimeter.Two separate ML models were developed for estimating the dose in terms of H p (10) and H p (0.07).The study shows that excellent accuracy can be achieved in the estimation of H p (d) by using CaSO 4 :Dy TLD badge with the help of ML models.
Furthermore, the present study focuses on the introduction of a data-driven approach aimed at enhancing radiation protection for occupational workers.We have developed a novel multi-stage ML technique that integrates energy estimation and dose estimation, resulting in a substantial enhancement in the precision of dose assessments, effectively transcending the limitations inherent in conventional algorithms.This novel approach is poised to play a pivotal role in strengthening occupational worker and installations safety, particularly within the diagnostic, medical, and nuclear industries where use of ionizing radiation is involved.The present study also demonstrates the significance of ML towards improving the occupational health and safety in radiation environments.

TLD badge
The TLD badge has three TL elements (Disc), each of 13.3 mm diameter and 0.8 mm thickness.They are clipped to an Aluminum plate to form the TLD card, which is inserted into a cassette having three filters/regions for the TL discs.The top disc (D1) placed between a composite metal filter, consisting of an Al filter with a thickness of 0.6 mm and Cu filter with a thickness of 1 mm.The middle disc (D2) is provided with a 1.6 mm thick Polystyrene filter.The bottom disc (D3) does not have any filters; it is covered by a polythene pouch and a paper wrapper.Figure 1 illustrates the pictorial representation of the TLD card and cassette with the filters.It is to be noted that the clip for wearing/fixing TLD badge is not shown in the figure 1.

Calibration and readout of TLD badge
TLD cards are read in the semi-automatic TLD badge reader using clamped heating profile [24].The hot N2-gas is directed towards the TL element, maintaining a temperature of 285 • C for 30 s.The TL emitted during the heating cycle is recorded at every second as TL counts.For routine calibration of the TLD badge reader, TLD badges are exposed to Cs-137 gamma radiation at a distance of 50 cm in a panoramic geometry, and the badges are read in a TLD reader after 7 d to eliminate the effect of fading of low-temperature peaks.The reader calibration factor, which represents dose per unit TL counts is determined from the average TL counts from D1 and the delivered dose in terms of whole body dose or H p (10)/H p (0.07) [24].At present the personal dose is reported in terms of whole body dose and skin dose [24].It is noted that the correspondence between the air-kerma/exposure for the panoramic in-air irradiation and the whole body dose/H p (10) and H p (0.07) for on phantom collimated irradiation has already been established.
To address the crosstalk observed when the three TL elements in TLD cards are sequentially heated in the TLD badge reader [25], disc calibration factors are applied.The disc calibration factors are determined by irradiating the TLD card in a 3 mm thick Perspex build-up so that all TL elements receive the same dose.After the readout, TL counts from D1 are used as a reference, and the TL counts from D2 and D3 are scaled to match D1.The scaling factors, i.e. disc calibration factors, along with reader calibration factors are always applied during the readout of service TLD badges.

Dataset preparation and preprocessing
Response of TLD badge in terms of operational quantities has been reported in a few studies [9,10,12,13,26].These experimental TL response datasets were compiled, consisting of approximately 700 sets of TL readouts.The dataset contains response to various photon beams of energy 12.4 keV-1.25 MeV, angles of incidence 0 • ± 60 • [26][27][28][29], mixture of different photon beams delivered H p (10)/H p (0.07) in the range of 0.21-440 mSv.In addition to the experimental data, TL counts per unit delivered H p (10) and H p (0.07) were utilized to simulate around 3700 sets of TL counts for mixed photon exposure.These simulations were performed considering different photon beam qualities and angles of incidence.To generate a feature set for the training of ML models, ratios of the TL counts from three elements were taken.These ratios were chosen to characterize the dosimeter response to specific beam qualities and orientations.

ML algorithms
Two ML algorithms namely Random Forest (RF) and ANN [30,31] were utilized in the present study.The RF is an ensemble learning algorithm that works by constructing multiple decision trees (ntree) and each tree in the forest is grown by considering a random subset of features (mtry), which helps to reduce over fitting.ANN, on the other hand, is a feed-forward neural network that consists of an input layer, one or more hidden layers, and an output layer.Each layer is made up of one or more neurons that use an activation function to transform the input signal and adds non-linearity.During training, the network adjusts the weights of the connections between neurons using back-propagation, which is a gradient-based optimization algorithm that minimizes the difference between the predicted and actual values.The performance of an RF and ANN model is highly dependent on the choice of hyper-parameters.Therefore, the ML models were trained using different sets of hyper-parameters selected systematically using the grid search approach, and the model showing the best performance for the validation dataset was selected.The ANN and RF models are developed using the R programming language and the related packages [32][33][34][35][36][37].
In the present study, the TL counts and ratios of TL counts from elements under different filters were utilized as inputs for the ML model, and the response variable, which relates the TL counts to H p (d), is estimated as an output of the ML models.In the first stage of the algorithm, i.e. prediction of photon energy into broad range (i.e.classes), the RF algorithm was utilized.For the estimation of average energy (Stage 2), both RF and ANN were used and compared.In Stage 3, the RF model is used for the estimation of H p (d).At Stage 2, the ratios of TL count along with the class of energy predicted by Stage 1 serve as an input to the ML model for the estimation of the logarithm of average energy of photons.At Stage 3, the class of energy predicted at Stage 1, average energy predicted by Stage 2, and ratios of TL counts were used as input to the ML model for estimation of operational quantity.
The selection of the RF model for estimating personal dose equivalent H p (d) is motivated by several key considerations, particularly the challenges posed by the normalization of the data.The range of variation in the ratio of TL counts spans from 200 to 0.75 in certain cases and 0.8-1.25 in others.Similarly, the values of average energy exhibit significant diversity, ranging from 12.4 keV to 1250 keV.Attempting to normalize the data under these circumstances can lead to very small feature values, resulting in poor differentiation of characteristics.This, in turn, may severely impact the performance of the ML model.In light of these challenges, the RF model presents a compelling solution.RF models are well-suited to handle diverse datasets with varying feature ranges without the need for extensive normalization [32].Unlike some other ML models that heavily rely on data normalization, RF models are robust and effective even when confronted with significant variations in feature scales.

Flow chart of the algorithm
To provide an overview of how the algorithm functions, a flow chart (figure 2) is created to depict the key stages and processes involved.The algorithm utilizes the net TL counts, specifically N D1 , N D2 and N D3 , which correspond to TL elements D1, D2, and D3 respectively.These net TL counts are utilized to compute various ratios.In Stage 1, these ratios are then fed into an ML model to determine the class of photon energy.In Stage 2, the class information, along with the ratios, is inputted into another ML model to estimate the logarithm of average energy.In Stage 3, the average energy, in conjunction with the ratio of TL readings, is employed to estimate the operational quantity H p (d) through regression models that yield a correction factor (CF) associated with H p (10) and H p (0.07).

Implementation in routine monitoring
In our ongoing pursuit to enhance dosimetric precision and operational efficiency within the routine TLD personnel monitoring, we have initiated the implementation of an ML-based GC screening algorithm.The GC screening ML models has been incorporated into a user-friendly graphical interface created through the 'R-Shiny' application [38].The present research plays a vital role in advancing this initiative, contributing to the improvement of a more robust and precise dose estimation module within a system intended to optimize and streamline operations in TLD personnel monitoring laboratories.
It is noteworthy that the ML models developed in this study do not require retraining for each batch of TLD dosimeters, as the batch-to-batch sensitivity variations are taken care by TLD reader calibration process.Also, employing new batch of dosimeters in the service does not require re-characterization of energy/angular response.

Energy dependence of TLD badge
The energy dependence of TL dosimeter was studied using S-Cs beam quality as a reference.Figure 3 and table 1 show the angle-averaged relative response (R Hp(d) ) of the TLD badge.In figure 3(a), for disc D3, R Hp (10) is approximately 100 at 12.4 keV due to high photoelectric interaction.As photon energy increases, R Hp (10) decreases from 13 to 4 (30-80 keV) and further to 1.7 (80-200 keV), reaching approximately 1 beyond 200 keV due to decreasing photoelectric effect and increasing Compton scattering.For disc D1, at lower energies, both R Hp (10) and R Hp(0.07) are low due to attenuation offered by the composite metal filter.R Hp (10)   increases until 80 keV due to increased penetration and higher photoelectric cross section of dosimeter at this energy range, then decreases gradually with higher energies due to increasing Compton cross-section.Similarly, figure 3(b) shows a comparable behavior of R Hp(0.07) , with the only difference being an the over-response of ∼10 for disc D3 at lower energies.This over response difference is primarily due to the delivery of the same amount of air kerma, which results in significantly higher H p (0.07) compared to H p (10) at the lower photon energies.

Angular dependence of TLD badge
The TLD badge's angular response, in addition to its energy dependence, is influenced by disc geometry, filter arrangement, and air kerma-to-personal dose equivalent conversion coefficient variations [26].Figure 4(a)'s polar plot illustrates the angular dependence at N-80 photon beam quality.Disc D1's response decreases with increasing angle due to x-ray attenuation by metal filters.In contrast, D2 and D3 show an increasing response, probably due to increased energy deposition with longer path lengths for x-rays inside the dosimeter.Filter thickness effects are negligible at higher photon energies.Figure 4(b) displays the angular response at S-Co beam quality.

Nature of variation of conversion coefficients and dosimeter response
Conversion coefficients (h pK ) are employed for deriving H p(d) from air-kerma in external photon radiation exposure.In figure 5, surface plots depict the variation of h pK (10) and h pK (0.07) with photon energies (E) and angles of incidence (α).Both (a) and (b) in figure 5 show similar behavior, with h pK (10; E, α) slab exhibiting slightly more angular dependence [27], particularly at lower photon energies.Ideally, a dosimeter mirroring h pK (d; E, α) slab response across energy and angle would require no additional correction for H p (d) estimation.However, practical deviations (figures 6 and 7) arise due to factors like dosimeter geometry, filter material, effective atomic number, necessitating algorithms with mathematical functions for mapping dosimeter element responses to H p (d).

Development and evaluation of ML models 3.4.1. Classification of energy
As a first stage of dose estimation algorithm, the classification model is developed to categorize the photon energies in to four classes.Consequently, the dataset was partitioned into four energy classes, ensuring equitable representation in each: Class 1 (E ⩾ 200 keV), Class 2 (100 keV ⩽ E < 200 keV), Class 3 (46 keV ⩽ E < 100 keV), and Class 4 (E < 46 keV), with an approximately equal number of samples in each category.
The classification model was developed using RF algorithm.It is a well-established fact that the number of features utilized by each tree is known to influence the accuracy of individual trees [32].As the mtry increases the accuracy of each tree increase.However, the correlation between the trees also increases with mtry [32].Typically, a rule of thumb is to set mtry to the square root of the total number of features (i.e.mtry = √ 8 ≈ 3).This approach effectively mitigates concerns related to inter-tree correlation.Subsequently, the model was tuned for optimum number of trees using repeated k-fold cross-validation and grid search for the ntree parameter, while keeping mtry constant.The training dataset comprises of 80% of   randomly selected data points from whole dataset of 745 data The mean accuracy from tuning shown is figure 8 shows no substantial improvements beyond ntree = 50.Hence, the RF model with mtry = 3 and ntree = 50 was deemed optimal.Furthermore, the performance of the RF model was evaluated using a separate test dataset, representing the remaining 20% of the overall dataset.A confusion matrix shown in table 2 was constructed to assess the model's performance on this independent test dataset.
Based on the confusion matrix the overall classification accuracy obtained for test dataset was 0.9799 (97.99%).Other performance parameters such as specificity, sensitivity, positive predictive value (PPV), negative predictive value (NPV), balanced accuracy, etc. are computed using a confusion matrix and tabulated in table 2 Sensitivity and Specificity gauge correct identification of TP and TN cases, while PPV and NPV assess predicted positive and negative cases.Detection prevalence tracks TP cases in the population tested, and balanced accuracy is the mean of sensitivity and specificity.The numbers from table 2 and the values of parameters in table 3, show that good classification accuracy is achieved using the RF model for the test dataset.

Estimation of average energy
Looking at the wide range of feature values, specifically in relation to the ratios of disc readings across a range photon energy, minor variations in these features are prone to being disregarded, consequently influencing accuracy.Considering this fact, prior to estimation of average energy the classification model was implemented to partition the data into discrete energy categories.This approach may serve to streamline the intricacies of the data and establish a robust framework, a crucial aspect in the context of radiation dosimetry.
Optimal model performance was achieved through hyper-parameter tuning using repeated k-fold cross-validation on the same dataset employed for the classification model.The ANN model, utilizing a sigmoid activation function, three hidden layers (8, 5, and 3 nodes), mean squared error loss, and self-adaptive gradient descent, exhibited the optimal predictive performance.The RF model, comprising 100 trees and 4 variables per split, also delivered good results, both models show R 2 values exceeding 0.95 with low mean absolute error (MAE) and root mean square error (RMSE).Comparing the models, the RF model outperformed the ANN, as seen in figures 9 and 10.The RF model displays slightly precise predicted-to-actual average photon energy ratios.It is to be noted that, this comparison does not serve as an absolute benchmark, as model performance may vary based on tuning parameters.
To explore the feasibility of a regression model independent of the classification model, both the ANN and RF models were tuned using the same approach and dataset.The performance metrics in table 4 closely   resembled those of models incorporating classification features.In practical situations, it is important to recognize that uncertainties can lead to variations in features.When dealing with datasets having few features, creating new features that capture key information from the existing ones becomes essential to enhance the model's reliability.These new features act as safeguards against noisy data, improve stability, and aid in data interpretation.Therefore, in present study, we have opted for a three-stage approach, but it is worth noting that a two-stage model remains a viable alternative.

Estimation of H p (d) 3.4.3.1. Preprocessing the data
In the second stage, the average photon energy was estimated.Progressing to the third stage, initially the correlation between the TL counts and personal dose equivalent was established by CFs.These CFs are designed to be dose-magnitude-independent, making them well-suited for ML models due to finite range.The derivation of CFs is based on a judicious weighting of TL counts from three TL elements, taking into account their significance in the context of the operational quantity.The CFs are computed using following empirical formulas: where, N D1 , N D2, and N D3 are net TL counts in µSv from elements D1, D2, and D3 respectively, while ρ, φ, and ω are constant weightage with values 0.2 µSv −1 , 0.4 µSv −1 , and 0.2 µSv −1 respectively.Additionally, The selection of coefficients such as β, ρ, φ, and ω was driven by their relevance to the TL counts from specific discs in relation to the operational quantity.When computing CF 1 , β's behavior becomes apparent: it displays an increasing trend with higher photon energy, thereby assigning greater weight to TL counts from D1.In contrast, at lower energies, β decreases, diminishing the contribution of TL counts from D1, reflecting their diminished importance at lower photon energies.However, the story changes when we consider TL counts from D2 and D3, which exhibit an over-response varying from 10 to 4, yielding substantially higher values.This observation validates our choice of a weighting factor, ρ = 1/5 µSv −1 , and underscores its effectiveness in ensuring a optimum correlation between CF 1 and H p (10).
Likewise, for CF 2 , the value of φ is chosen as 0.4, providing approximately 80% weightage to TL counts from D2 and D3 in the calculation of CF for H p (0.07).This choice is based on the nature of R Hp(0.07) depicted in figure 7(b), which demonstrates a reasonably good correlation between TL counts from D2 and H p (0.07).In contrast, the value of ω (representing the weightage of TL counts from D1) is kept at a minimum value of 0.2, as indicated by the R Hp(0.07) of D1, which suggests that TL counts from D1 are not reliable at lower photon energies.

Training and evaluation of regression model
In the stage 3, for training and tuning of ML model, the input parameters included energy class, average photon energy, TL counts, and their ratios, with the output being the corresponding CF.The RF model was tuned to have ntree = 100 and mtry = 4, achieving optimum performance as assessed by MAE and R 2 values, which are presented in table 5.A correlation plot in figures 11(a) and (b) illustrates the relationship between predicted and actual CF for H p (10) and H p (0.07), respectively.
Additionally, figures 12(a) and (b) presents histograms of the ratio of predicted and actual CF.Notably, the majority of predicted CF values fall within ±25% of the actual CF, indicating a reasonably accurate estimation.Moreover, most ratios in figure 10 are within the range of 0.9-1.1,demonstrating good agreement between the ML-predicted CFs and the actual values for both H p (10) and H p (0.07).

Comparison with existing algorithms
Presently, a linear combination-based algorithm for the estimation of H p (10), and a polynomial-based algorithm for the estimation of H p (0.07) are recommended [8].These algorithms were developed and further refined by a task group at BARC.For more details on conventional approach, one may refer to the study by Pradhan et al [8].Brief details of the algorithms are given in the table 6.
A comparison of ML models and existing algorithms was performed using a dataset comprising 125 and 144 sets of TL readouts for H p (10) and H p (0.07) estimation, respectively.This dataset covers various photon beam qualities (N-15 to S-Co) and known delivered dose (H ref ).The results are presented in figures 13 and 14, demonstrate the improved performance of the ML algorithm over existing algorithm.Additionally, table 7 indicates a 7.2% reduction in the coefficient of variation (CoV) for H p (10) and a 4.3% reduction in the CoV for H p (0.07), suggesting that approximately 95% of the estimates will fall within ±22.4% of delivered H p (10) and ±25.2% of delivered H p (0.07).These values align with acceptance criteria and performance limits for personnel dosimeters [24].

. Sources of uncertainty
The uncertainty in a 4 :Dy-based TLD badge system is categorized into two distinct types: random (Type I) and systematic (Type II) uncertainties.Random uncertainties can theoretically be reduced through an increase in the number of measurements and stem from factors such as inhomogeneity in detector sensitivity and fluctuations in detector readings at zero dose.In contrast, systematic uncertainties,   often labeled Type II, persist regardless of repeated measurements and include factors like energy dependence, directional dependence, non-linearity of the response, fading dependent on ambient temperature and humidity, effects due to exposure to light, effects resulting from exposure to types of ionizing radiation not intended for measurement etc.Out of above the energy and directional dependence are considered as a major contributor to the total uncertainty.For more detailed information on this topic, please refer to the source [39] listed in the references.

Limitations of the model
In the case of a few simulated mixed-field exposures involving low-energy photons and S-Cs or S-Co photons, we observed a significant underestimation of H p (10) doses, reaching up to 40%.This underestimation was particularly pronounced when there was a combination of photon beams below 40 keV and S-Cs beams, resulting in TL count patterns resembling exposure to photons less than 80 keV. the models generated CFs appropriate for ∼80 keV photons, to the underestimation of doses.It is important to note that in such scenarios, existing algorithms also demonstrated a underestimation of H p (10).This underestimation can be attributed to the strong energy dependence of the TLD badge response at lower photon energies.However, it is crucial to emphasize that these exposure combinations are highly unlikely to occur in typical workplace settings.
It is to be noted that in case of the occupational exposure in field conditions, it is imperative that the TLD be getting the exposed from all the directions.Therefore, the angle of exposure has not been estimated using ML-models.However, the while developing ML models, these models were trained on the dataset which consist angular exposures, making them capable of accounting for it.

Conclusion
In view of the substantial variability arising from energy and angular dependence, the application of ML techniques for occupational dose estimation shows great promise.Furthermore, the present study introduces a novel method for precise estimation of average photon energy using only three TL elements from a routine TLD badge.By incorporating photon energy class and average photon energy as supplementary factors in our ML models, impressive accuracy in the estimation of operational quantity is obtained.The ability to accurately estimate H p (d) based on TL data from existing TLD badges has practical implications for occupational radiation monitoring and ensures reliable and precise dose assessments.
Moving forward, future work aims to expand the capabilities of ML models by incorporating the estimation of H p (d) for beta and beta photon exposures.This extension will further enhance the applicability and utility of our approach in a wider range of occupational radiation scenarios.

Figure 1 .
Figure 1.Diagrammatic representation of the TLD badge, showcasing the configuration of filters within the TLD cassette [23].

Figure 2 .
Figure 2. Flow chart illustrating the multi-stage ML model for estimation of personnel dose equivalent.The boxes with light blue background color represent stages of the algorithm.

Figure 3 .
Figure 3. Normalized photon energy-dependent response of TLD badge elements concerning personal dose equivalent, referenced to the response at S-Cs beam quality.

Figure 4 .
Figure 4. Angular response characteristics of CaSO4:Dy-based TLD badge to N-80 and S-Co beam qualities.

Figure 5 .
Figure 5. Depiction of air-kerma to personal dose equivalent conversion coefficients, as per ISO-4037-3, in relation to the average energy of photons (E) and the angle of incidence of radiation (α) [27].

Figure 6 .
Figure 6.Surface plots illustrating the variation in TL discs response with respect to delivered Hp(10), when exposed on ISO-slab water phantom to average photon energy (E) ranging from 30 keV to 1250 keV and angle of incidence (α) from −60 • to +60 • .

Figure 7 .
Figure 7. Surface plots illustrating the variation in TL discs response with respect to delivered Hp(0.07), when exposed on ISO-slab water phantom to average photon energy (E) ranging from 12.4 keV to 1250 keV and angle of incidence (α) from −60 • to +60 • .

Figure 8 .
Figure 8. Plot illustrating the mean accuracy values obtained during the tuning of ntree using 10-fold cross validation approach with number of repeats = 30.

Figure 9 .
Figure 9. Correlation plot depicting the relationship between the actual average photon energy and the predicted average photon energy using random forest and artificial neural network models.

Figure 10 .
Figure 10.Histogram of the ratio of predicted average photon energy to actual photon energy, facilitating a quantitative evaluation of the performance of random forest and artificial neural network models.

Figure 11 .
Figure 11.Correlation plots of machine learning-estimated correction factor and the actual correction factor values for Hp(10) and Hp(0.07).

Figure 12 .Table 6 .
Figure 12.Histogram of the ratio of predicted correction factors to actual correction factors, facilitating a quantitative evaluation of the performance of random forest in estimating the operational quantities.

Figure 13 .
Figure 13.Comparison of relative deviations of estimated personal dose equivalent using existing algorithm and machine learning algorithm.

Figure 14 .
Figure 14.Boxplots showing the distribution of relative deviation in estimated personal dose equivalents using existing algorithms and machine learning-based algorithm.

Table 1 .
Average relative response of TLD badge with respect to the delivered personal dose equivalent.
Radiation QualityAverage Energy (keV)TL counts per delivered operational quantity (R Hp(d) )

Table 2 .
Confusion matrix of evaluation of classification model using test dataset.

Table 3 .
Performance evaluation of the classification model for test dataset.

Table 4 .
Performance evaluation of regression models for estimation of energy.

Table 5 .
Performance evaluation of random forest based regression models for estimation of correction factors for estimating operational quantities.

Table 7 .
Comparison of coefficient of variation of relative deviations of the estimated doses by existing algorithms and machine learning algorithm.