Estimating Compton scatter distributions with a regressional neural network for use in a real-time staff dose management system for fluoroscopic procedures

Abstract

Staff-dose management in fluoroscopic procedures is a continuing concern due to insufficient awareness of radiation dose levels. To maintain dose as low as reasonably achievable (ALARA), we have developed a software system capable of monitoring the procedure room scattered radiation and the dose to staff members in real-time during fluoroscopic procedures. The scattered-radiation display system (SDS) acquires imaging-system signal inputs to update technique and geometric parameters used to provide a color-coded mapping of room scatter. We have calculated a discrete look-up-table (LUT) of scatter distributions using Monte-Carlo (MC) software and developed an interpolation technique for the multiple parameters known to alter the spatial shape of the distribution. However, the file size for the LUT’s can be large (~2GB), leading to long SDS installation times in the clinic. Instead, this work investigated the speed and accuracy of a regressional neural network (RNN) that we developed for predicting the scatter distribution from imaging-system inputs without the need for the LUT and interpolation. This method greatly reduces installation time while maintaining real-time performance. Results using error maps derived from the structural similarity index indicate high visual accuracy of predicted matrices when compared to the MC-calculated distributions. Dose error is also acceptable with a matrix element-averaged percent error of 31%. This dose-monitoring system for staff members can lead to improved radiation safety due to immediate visual feedback of high-dose regions in the room during the procedure as well as enhanced reporting of individual doses post-procedure.

1. INTRODUCTION

Fluoroscopically-guided interventional procedures may lead to long exposure times. These procedures require care by the interventional staff members to not exceed thresholds for latent skin effects, as well as keeping in mind stochastic risks, while also maintaining sufficient image quality and care. To reduce risk to the patient, our group has developed the Dose Tracking System (DTS), which acquires digitized machine parameters in real-time via access to a fluoroscopic machine’s application programming interface (API).¹ The DTS is currently a product distributed with various clinical fluoroscopic systems by Canon Medical Systems Inc.

In addition to patient dose monitoring during procedures, various sources imply a need for minimization of staff member doses in accordance with the as low as reasonably achievable (ALARA) paradigm. The NCRP and ACR, as a couple of examples, have begun improving the standards for fluoroscopic operator and staff education, but neither efficiently educate on optimal locations in the room with regards to scatter dose.^2,3 Further, studies have indicated that interventionalists may overuse high dose-rate techniques, such as digital subtraction angiography (DSA), and may neglect the use of protective apparatus such as leaded eyewear and ceiling-mounted shields.⁴

A paradigm shift has thus begun to emerge where several groups have are developing safety training systems using augmented reality (AR) and virtual reality (VR) to assist staff members and residents in their understanding of dose received, and potential changes in practice to avoid high-dose regions in the spatial distribution of scatter as well as high dose-rate techniques when available. One such system is XAware-Live, which makes use of a high-end graphics processing unit (GPU) to accomplish fast Monte Carlo (MC) simulations with greater than 10% error.⁵ To supplement the training, the group which has developed XAware-Live has integrated the use of real-time dosimeters into the system. This additional feedback facilitates dose monitoring over time for each staff member in order to identify moments of high dose for evaluation in the training system.

Using VR for training system development, two groups have been contributing to the paradigm shift: one based in the University of Science and Technology of China, and one based in Flensburg University in Germany. One of the groups, based in the University of Science and Technology of China, incorporated the Microsoft Hololens for viewing and interacting with the training system.⁶ This simplistic system allows for the trainee to alter geometric and exposure parameters, but doesn’t yet involve a function for post-procedural analysis. The second system (being developed at Flensburg University) includes the option to view 3D scatter distributions, computing dose with a virtual probe, as well as sampling two-dimensional planes of scatter.⁷ However, similar to the other system mentioned, there is no extensive method available for post-procedural dose analysis. Both of these groups rely upon high-end GPUs for on-the-fly MC calculations.

In contrast to other groups, we have primarily focused on the development of a real-time staff dose monitoring software, called the Scattered Radiation Display System (SDS).⁸ This system, after completion of development, can be adopted as an additional component of the DTS to be installed in various interventional radiology (IR) clinics. Real-time performance of our system, as opposed to methods implementing expensive GPUs, is accomplished through the use of a pre-loaded dose matrix library. EGSnrc MC simulations are used for generating these matrices with a very large number of photon histories (1 billion), facilitating sub-10% error, depending on our choice of interpolation coarseness. The SDS’ design is compatible with any clinical system incorporating the use of an application programming interface (API), such as the one we have been testing on, the Toshiba Biplane Infinix-i. As such, we have a system which can function on a wide variety of computers at a low cost while still providing high fidelity and accuracy, as well as highly detailed post-procedural dose analysis.

A limitation of the dose matrix library is storage. Our goal is to develop a final system which incorporates three-dimensional distributions. Distributions we have produced can take up to 2 GB each, which can then lead to very long software load times and installation times (discussed later in this report). We have thus designed a regressional deep neural network (RNN), which acquires unique parameters known to alter the spatial shape of scatter distributions as input in order to predict a scatter distribution. In this work we investigate the viability of this alternative method for scatter computation, to improve the functionality of the SDS by eliminating long installation and load times.

2. MATERIALS AND METHODS

2.1. SDS Architecture, Scatter Distribution Library, and Dose Rate Algorithm

Our group has designed the SDS to provide staff members with a virtual, top-down representation of the interventional room with a superimposed scatter distribution. The SDS obtains machine parameters in real-time using a controller area network (CAN) bus interface. These parameters consist of tube voltage, tube current, collimator position, patient table position, gantry orientation, and beam filter type to name a few. Once the parameters’ values are derived, they are fed into the dose rate algorithm in Equation 1:

$$M_{\mathit{\text{Final}}}\left(\frac{mGy}{hr}\right)=M_{\mathit{\text{Interpolated}}}\left(\frac{mGy\mathit{\text{scatter}}}{Gy\mathit{\text{entrance air kerma}}}\right)*DT{S}_{cal}\left(\frac{mR}{mAs}\right)*I(mA)*ISL*C\left(\frac{mGy}{mR}\right)*3.6\left(\frac{s*Gy}{hr*mGy}\right)$$ (1)

where M_final is the current scatter distribution matrix whose elements have units of mGy/hr, M_interpolated is an interpolated scatter matrix derived from discrete matrices stored in the LUT given per unit entrance air kerma, DTS_cal is a calibration factor measured at the interventional reference point (IRP), I is the tube current in units of mA (for pulsed fluoroscopy, this is modified by the duty cycle), ISL is an inverse square law correction for skin location changes from the IRP, C is a scalar correction to convert from units of exposure to air kerma, and the final term is an additional unit correction factor adjusting time and air kerma.

For parameters known to alter the spatial shape of the patient’s scatter distributions, we have produced an LUT at discrete parameter values using EGSnrc Monte Carlo (MC) for access by the system. The resulting distribution is then color-coded and overlayed on the virtual background (see Figure 1). The virtual background consists of C-Arm gantry, patient table, and patient graphic models. The patient model (created using MakeHuman⁹) is rendered through the use of an STL read function, which obtains spatial information from files storing the vertices and faces. The Zubal anthropomorphic phantom is employed in our MC simulations for estimating scatter from a human of average size; however, the scatter distribution will be somewhat dependent on patient size and shape and we will be looking at these dependencies with different computational phantoms in other studies.

Fig. 1:

An example SDS image frame displaying a top-down, virtual representation of the interventional room. Scatter distributions are color-coded and overlayed on the virtual background while dose is presented at the staff member’s location using a black, circular position/dose rate indicator. An example of person shielding is included in this image where the trapezoidal region beyond the staff member indicator delineates low scatter dose values due to attenuation. Examples of object shielding are not shown in this demo image for simplicity.^8,10,11 [Note: top-down means viewing the patient on the table from the ceiling]

2.2. Regressional Deep Neural Network Architecture, Training, and Metrics for Performance Evaluation

In the current work, we seek to obtain a first-order prediction of the scatter distribution, thus differences due to shielding are corrected for using methods we have developed previously for people and objects in the room. Therefore, when discussing scatter dose accuracy later in this report, we are referring to first-order approximations by our RNN, which would then be corrected using our shielding algorithms.¹⁰

For comparison in this study, we have made use of 2D scatter distribution matrices, as opposed to the 3D matrices we are currently developing for other applications. The plane chosen is at a height of 1 m above the floor, which approximately corresponds to gantry isocenter height as well as a typical waist height for an individual.

Figure 2 illustrates the interpolation scheme we have employed due to the discrete nature of the values in the LUT. That is, when machine parameters lie in an interval between two values in the LUT, interpolation is required otherwise an unacceptably large LUT would be required with every possible combimation of parameters.

Fig 2:

Graphic representation of the LUT for scatter dose values shown as the discrete data points on the curve as a function of RAO-LAO angle. All dose values in the figure correspond to a single voxel location in the scatter dose matrix. Similar tables are stored for other parameters known to alter the spatial shape of the distribution, such as tube voltage and entrance field size. This method requires interpolation between dose values for intermediate data points not in the LUT so as to acquire a continuity of dose tracking during procedures. [Note: we are currently employing linear interpolation between data points]

In the current work, we focus on the development and evaluation of an RNN we have developed in TensorFlow for replacing the LUT. The RNN functions by first obtaining an input vector which consists of the five parameters known to alter the spatial shape of the distribution (half-value layer {combined tube voltage and beam filter}, entrance field area, patient table shift, RAO-LAO gantry angle, and CRA-CAU gantry angle). Our RNN architecture consists of 3 dense layers, and begins with the aforementioned input vector. The output layer then consists of a vector which is reshaped to obtain the scatter distribution matrix. 30,000 epochs were used to achieve a large enough validation accuracy for prediction of very low doses (order of 10⁻³ $\frac{mGy}{Gy}$) to occur. In addition, a training, validation, testing split of 65%, 20%, and 15% was used on a total of 250 distributions. Figure 3 displays the RNN architecture. The previously mentioned dataset for the RNN is composed of ground truth data (distributions) developed using DOSxyznrc with particular variations in the spatial shape altering parameters.

Fig 3:

The RNN architecture our group designed for predicting scatter matrices. With an RNN, we seek to solve a regressional problem where a numerical input is mapped to a numerical output. For our work, the input layer consists of a 1×5 vector whose elements contain values for the five parameters mentioned in the text. To satisfy the regressional problem, a mean squared error loss function was implemented in addition to three hidden layers with rectified linear (relu) activation functions, as well as a relu activation prior to the output mapping (for suppressing negative dose predictions). The output in this case is a 1 × 202500 vector whose elements contain scatter dose values. This vector is then reshaped into a 450 × 450 scatter matrix. A k-fold cross-validation scheme, with 10 folds, was employed so as to avoid bias in the model validation.

To compute the structural accuracy of the RNN predictions, we have derived an error map from a map of the structural similarity index¹², defined in Equations 2 and 3 below:

$$\begin{gathered} \mathit{\text{SSIM}}(x,y)=[l(x,y)][c(x,y)][s(x,y)] \\ -1\le \mathit{\text{SSIM}}\le 1 \end{gathered}$$ (2)

$$E=100*\left[I-\frac{1}{2}\left(SSI{M}_{map}+I\right)\right]$$ (3)

where SSIM(x, y) is the structural similarity index in a given image pixel defined in terms of l(x, y) (luminance), c(x, y) (contrast), and s(x, y) (relative pixel offset). Given that SSIM values range from −1 to 1, the structural error map matrix, E, was derived by generalizing this equality using matrices where SSIM_map is a matrix of SSIM(x, y) values and I is an identity matrix with the same dimensionality as SSIM_map. [Note: the matrix E is identical to a map of the structural dissimilarity index DSSIM] In addition, we made use of Hadamard division to derive a percent error map for expressing the accuracy of dose rate predictions, as shown in Equation 4:

$$PE=\mid \mathit{\text{Prediction}}-MC\mid \oslash MC$$ (4)

where PE is the percent difference map, Prediction is the scatter matrix predicted by the RNN, and MC is the ground-truth Monte Carlo matrix. Further, the time of computation using the RNN was compared to the distribution load time in the SDS using the LUT as well as to the time for interpolation using Matlab and Python’s built in stopwatch functions.

1. RESULTS

3.1. Structural and Dose Error Evaluation

Figure 4 presents structural accuracy and percent difference results for the scatter distribution map (scatter emanating from the chest region) with a CRA angulation of 5 deg. at 80 kVp (1.8 mmAl added filter), as well as a patient table shift of 5 cm to the left of isocenter from the top-down perspective. As can be observed in Figure 4C, the structural accuracy is good in regions where staff members will typically be located during procedures (>50 cm radius from beam isocenter (image center)). However, when comparing to Figures 4A and 4B, it can be observed that the network performs less well in high dose gradient regions. It is also apparent from the results in Figure 4D that, overall, dose prediction accuracy is acceptable with an average percent difference of less than 20%. While some regions contain higher percent differences, they mainly lie within a 50 cm radius from isocenter where staff members will not be located.

Figure 4:

A) RNN prediction of a scatter distribution for 80 kVp, 1.8 mm Al beam filter, 5 cm patient table shift left of isocenter, and a 5 deg. CRA gantry tilt. B) MC (ground truth) results for the same configuration as A. C) Structural error (DSSIM) map derived from the structural similarity index, SSIM, discussed in equations 2 and 3. We can see that there is very good agreement (<10% structural error) for most of the pixels in the images. Regions where the dose gradient is the largest appear to have the largest error (40%). D) Percent absolute difference map indicating the dose prediction accuracy in each matrix element. While there are regions of high error (> 20%), these tend to lie within 50 cm of isocenter. Further, the average dose error is reasonably low (< 20%) across the whole matrix and in regions where staff members will typically be located (> 50 cm). [Note 1: the horizontal and vertical axis labels in the images express room coordinates (units of cm) with beam isocenter at the origin. Note 2: the pixel dimensions are 0.4 cm × 0.4 cm.]

Figure 5 is an overview of network performance on the testing dataset and gives a plot of values of the structural error averaged over all matrix elements (element-averaging) in each distribution map. The majority of scatter matrices in the testing dataset had an element-averaged error below 15%, which we deem good performance. This implies that staff members can acquire reliable visual feedback of the scatter distribution structure and, hence, recognition of high dose rate regions using the network.

Figure 5:

In this figure is an overview of results for the scatter matrix element-averaged structural error. The horizontal axis expresses the index of each matrix in the testing set used to compare against the RNN-derived matrix for the corresponding configuration; the order of the index is semi-random. We can see that most of the distributions (57%) have a good structural agreement (less than 15% error) and 42% of the distributions have a fair agreement (between 15% and 30% error). Only three predictions (1%) performed poorly, with an element-averaged structural error greater than 30%.

In addition to structural error, we analyzed the element-average absolute percent error over the testing cohort. In Figure 6 we see that approximately 67% of predictions performed favorably, with an error below 30%. Analyzing this data in more detail (as can be seen in Figure 7C for a dose prediction with an element-averaged percent error of approximately 50%) indicates that much of the high error lies in regions containing very low scatter dose, or regions where the primary beam is located. We quantified the high dose region as values lying above a threshold of 0.1 mGy/Gy in the neural network predicted matrix. Figure 8 demonstrates that within this region the element-averaged absolute percent error is less than 20% for all of the distributions in the testing cohort, except for 5. Thus, some of these high percent error averages are misleading given that feedback of reasonable error can be maintained at typical staff locations in a range of 50 cm to 100 cm from beam isocenter, or in regions of concern where dose is higher. However, two predictions by the network led to complete failures with percent errors of 88% and 140% averaged across the entire matrix, as well as 81% and 40% in the high dose region, respectively. These particular failures occurred for interpolations between parameters which are not as well represented in the dataset. Therefore, we believe that these very large errors, as well as errors in low dose regions far from the source or opposing the direction of the beam, can be mitigated further by expanding the dataset as well as suppressing details from the primary beam. We are currently limited by the number of distributions (250), so, naturally, by filling in the missing information we should see a reduction in failed tests. Further, we also believe that a weighting scheme may be applicable such that smaller penalties are applied where quantum fluctuations are high.

Figure 6:

In this figure is an overview of results for the scatter matrix element-averaged absolute percent error. The horizontal axis expresses the index of each matrix in the testing set used to compare against the RNN-derived matrix for the corresponding configuration; the order of the index is semi-random. 67% of the predictions demonstrated fair performance, with an element-averaged percent error below 30%. The other 33% performed poorly, however, as can be seen in Figure 7, these averages are sensitive to outliers, thus some of these predictions are acceptable for typical staff member locations. It should be noted that the errors of 140% and 88% are outliers, and represent cases of complete failure of the network for prediction.

Figure 7:

Sample results for an exposure using the following parameters: 70 degree LAO, 100 sq. cm entrance field area, 0 cm table shift, 80 kVp, and a 1.8 mmAl added filter. (A) The neural network prediction (B) The ground truth MC scatter distribution. (C) The percent error map comparing the doses elementwise. The element-averaged percent error was approximately 50%, while the element-averaged structural error was approximately 15%. As can be seen, the network performs poorly in regions where staff members would receive low dose. However, the network performs very well in regions of higher concern, where the dose is much larger. Additionally, while there are noticeable differences in the scatter distribution structure, they agree rather well. This implies that the staff members can still receive useful feedback in a scenario when the dose error is high. Further improvements should be made to reduce error in low dose regions so as to facilitate accurate individualized feedback in all scenarios. Still even with the high error, we believe this result is promising since the display provides useful feedback on the presence of scatter and its relative magnitude.

Figure 8:

In this figure is an overview of results for the scatter matrix element-averaged absolute percent error in the high dose region (> 0.1 mGy/Gy). The horizontal axis expresses the index of each matrix in the testing set used to compare against the RNN-derived matrix for the corresponding configuration, as discussed previously. Within this region the dose prediction fidelity is high, where 85% of the predictions lie within an average percent error of 20%. This indicates that within regions of higher concern, accurate individualized dose tracking can be maintained.

It should also be noted that during training a validation accuracy of 75% was achieved, indicating that our model is not overfit to the testing cohort. At 75% accuracy, we began to observe fair performance of the network, although some of the lowest values were being mapped to 0 mGy/Gy due to implementation of a relu activation function on the output. This underscores the need for further enhancement and adjustment of the network architecture and hyperparameters.

3.2. Evaluation of Neural Network Prediction and LUT Load Speeds

Latency due to RNN prediction time was compared to both LUT distribution load times as well as interpolation time in Figure 9. We note here that LUT distribution load times will be completely removed due to replacement by the RNN in the SDS. Further, it should be reiterated that for this study we are using 2D scatter matrices at a pre-selected height. In future work, we plan to incorporate the 3D distributions mentioned in the Methods section for other applications. This means that the 91 ms median LUT distribution load time discussed in Figure 9 will accumulate over the total number of scatter matrix elements in the third dimension as well as the total number of distributions in the LUT. Hence, per distribution there would be at least a 450-fold increase in load time, which could lead to software installation times of an hour or more. These results are impactful given that SDS installation and load times will be greatly improved, and real-time performance will not be lost.

Figure 9:

A boxplot comparison of the RNN prediction times to load times for each LUT distribution as well as to times for interpolation in the SDS. The median prediction time of 14 ms is significantly (P < 0.01) less than the median load time of 91 ms. The implication is that the LUT distribution load time will accumulate over the total number of distributions stored in the library, which will be removed completely when using the RNN. Further, the RNN prediction times will not be equivalent to loading in individual distributions. The median interpolation time of 3 ms is significantly (P < 0.01) less than the median RNN prediction time of 14 ms. However, the 14 ms time is sufficiently low for maintaining real-time performance.

4. CONCLUSIONS

The implication from the results in this study is that, using our RNN, staff members can obtain reasonably accurate visual feedback of high dose regions to avoid during the course of the procedures. Additionally, scatter dose value predictions, which are applicable to individualized dose recording (i.e. generation of dose reports for post-procedure review), are accurate in regions of higher concern (> 0.1 mGy/Gy). Accurate estimation of the dose per matrix element is pertinent in the context of reporting dose to different regions on the staff members’ bodies such as: the waist level, collar level, and eye lens level.¹³ The first step in accurate estimation of dose is the accuracy of the MC simulation and the ability to model the clinical situation. This is the goal of other studies. The second step, which is the focus of this work, is to determine if the RNN can accurately mimic the MC results. To improve on the agreement between the RNN and MC, we are currently in the process of expanding, enhancing network architecture, as well as developing methods to minimize the effects of the primary beam in the region where staff members will not be located (< 50 cm radius from isocenter). Currently, we are including the primary beam which generates very high values in the scatter distribution, which we believe contributes to higher errors in the neural network’s predictions. By removing such high values, the range of doses would be compressed, allowing for the predictions to converge more readily. With our current network, we have observed that, due to the implementation of a relu activation on the output layer, very small values are consistently mapped to 0 mGy/Gy. This error could also underscore the importance of longer training where the validation error may improve, subsequently implying better efficacy in elementwise predictions within low dose regions. Timing comparisons indicate that interpolation is more efficient from a real-time performance standpoint. However, real-time performance would not be lost upon implementation of the RNN and installation time, as well as load time, would be dramatically improved due to the removal of the LUT’s, which could have consisted of 2 GB sized distributions in their final state. A real-time system capable of monitoring staff member dose will improve radiation awareness and thus safety during the course of FGI procedures.