Machine learning-based prediction of MRI-induced power absorption in the tissue in patients with simplified deep brain stimulation lead models

Abstract

Interaction of an active electronic implant such as a deep brain stimulation (DBS) system and MRI RF fields can induce excessive tissue heating, limiting MRI accessibility. Efforts to quantify RF heating mostly rely on electromagnetic (EM) simulations to assess individualized specific absorption rate (SAR), but such simulations require extensive computational resources. Here, we investigate if a predictive model using machine learning (ML) can predict the local SAR in the tissue around tips of implanted leads from the distribution of the tangential component of the MRI incident electric field, E_tan. A dataset of 260 unique patient-derived and artificial DBS lead trajectories was constructed, and the 1 g-averaged SAR, 1gSAR_max, at the lead-tip during 1.5 T MRI was determined by EM simulations. E_tan values along each lead’s trajectory and the simulated SAR values were used to train and test the ML algorithm. The resulting predictions of the ML algorithm indicated that the distribution of E_tan could effectively predict 1gSAR_max at the DBS lead-tip (R = 0.82). Our results indicate that ML has the potential to provide a fast method for predicting MR-induced power absorption in the tissue around tips of implanted leads such as those in active electronic medical devices.

I. Introduction

THE rising prevalence of chronic diseases coupled with the rapidly aging population worldwide has made medical implants increasingly ubiquitous [1], [2]. More than 12 million Americans currently carry a form of orthopedic, cardiovascular, or neuromodulation device, and the market is projected to reach $160 billion by 2022 [1], [3]. Magnetic resonance imaging (MRI) is the diagnostic imaging modality of choice for the majority of neurological, cardiac, and musculoskeletal disorders. It is estimated that 50%−75% of patients with cardiovascular implants will need to undergo MRI during their lifetime [4], with many patients requiring repeated examinations [5]. Similarly, 70% of patients with neuromodulation devices such as those with deep brain stimulation (DBS) implants will need an MRI exam within 10 years of device implantation [6], either for ruling out complications or for optimizing stimulation protocols [7], [8].

Unfortunately, the interaction between MRI radiofrequency (RF) fields and conductive implants have led to fatal injuries due to RF heating of the tissue around the implant [9], making MRI inaccessible to such patients. In response, extensive effort has been dedicated to quantifying and mitigating the problem of MR-induced RF heating. These efforts heavily rely on electromagnetic (EM) simulations that represent details of human body heterogeneity [10], the implant’s structure [11], [12], and features of the MRI coil [13], [14]. As such, these simulations are notoriously time and memory consuming; even with today’s high-power computing clusters, a single simulation scenario can require hundreds of gigabytes of RAM and tens of hours to complete [15].

Machine learning (ML) has been recently suggested to estimate tissue properties and personalized SAR [16]–[18], and one group has applied it to assess MRI RF heating in the presence of passive implants [19], [20]. Specifically in [19], [20], a neural network was trained to predict the RF heating, as indicated by the specific absorption rate (SAR) of energy in the tissue around an implanted passive trauma treatment during MRI, given the geometric parameters of the implant. The algorithm was able to predict the maximum of 1g-averaged SAR around the implanted plate with high fidelity (correlation coefficient R = 0.97). However, the location and orientation of the implanted plate were fixed within the MRI coil; thus, the effect of variations in the implant’s exposure to the MRI electric field was not investigated.

In patients with elongated conductive implants, such as those with leads in active implantable medical devices (AIMDs), the distribution of the MRI incident electric field along the trajectory of the implanted lead is a decisive factor in the determination of RF heating [21]–[27]. A traditional approach for estimating MR-induced RF heating of elongated leads is through the concept of the lead’s transfer function. Here, the response of the lead to a stepwise tangential electric field that is applied locally along the length of the lead is calculated and used to estimate the response of the lead to an arbitrary incident electric field [28]. Based on this premise, we investigated whether a ML algorithm can be trained to directly map the MRI incident electric field along the trajectory of an implanted lead to the SAR amplification at the lead’s tip.

To examine the feasibility of the proposed technique, we adapted a ML algorithm to predict SAR amplification at the tips of simplified DBS lead models generated by a 1.5 T head coil. Our neural network focused on DBS leads with varying locations within the MRI coil. First, a dataset with 83 unique patient-derived DBS lead trajectories was constructed from postoperative computed tomography (CT) images of patients who underwent DBS surgery. The dataset was then augmented with an additional 177 artificial trajectories to account for intersurgeon variability in routing of the extracranial trajectory of DBS leads. A total of 260 full-wave EM simulations were performed to calculate the SAR amplification at the tip of each lead. We showed that a simplistic neural network (a feedforward, fully connected network with three hidden layers, rectified linear unit (ReLU) activation, and mean squared error (MSE) loss) can achieve successful SAR predictions. We performed hyperparameter optimization with five-fold cross-validation and trained the algorithm on 80% of the dataset to predict the maximum 1g-averaged SAR, referred to as 1gSAR_max, around the tips of lead models from the distribution of the tangential component of the MRI incident electric field, E_tan, sampled at 5 mm intervals along the lead’s trajectory. The remaining data (20%) were used as a test set where we demonstrate the potential of the ML algorithm to predict SAR. We showed the feasibility of ML-based predictions, indicating that an inexpensive, fast, and simple solution is possible for a rough estimation of RF-induced power absorption in the tissue around DBS leads in cases where full-wave EM simulations are too costly or time consuming to perform.

In what follows, we provide the methods for constructing the simplified DBS lead models, give details of EM simulations that calculate the ground-truth 1gSAR_max, and describe feature selection and construction of the ML algorithm. We then present the resulting architecture of the ML algorithm and its performance on predicting 1gSAR_max. Finally, we discuss the limitations and opportunities for future work.

II. Methods

A. Construction of Simplified DBS Lead Models

DBS lead models with patient-derived trajectories were constructed based on postoperative CT images of patients with DBS implants. Use of patients’ imaging data for the purpose of simulation and modeling was approved by Northwestern University’s Institutional Review Board. Fifty (50) patients aged 18–89 years old who underwent deep brain stimulation at Northwestern Memorial Hospital since 01–01-2014 were retrospectively recruited for the study. In total, 83 patient-derived DBS lead models with unique trajectories were reconstructed from medical images. DBS leads were identified in postoperative CT images using 3D Slicer 4.10.2 (http://slicer.org) visualization software. Once the lead artifacts were identified, preliminary 3D surfaces of lead trajectories were constructed, and the corresponding coordinates of points along the lead trajectories were extracted. A triangulated surface model of the patient’s head was also created from CT images and manually aligned to a standard homogeneous head model (σ = 0.49 /m, ε_r = 66.34) based on the Multimodal Imaging-Based Detailed Anatomical (MIDA) human head and neck model [29] via rigid transformation/registration (translation and scaling only, 6 degrees of freedom). The same transformation was then applied to the lead trajectories to allow for anatomically representative positioning of the leads in the homogeneous head model.

The extracted lead coordinates were used to reconstruct models of simplified DBS leads comprised of a conductive wire (σ = 4 × 10⁶S/m, diameter = 1 mm) embedded within a urethane insulation (σ = 0 S/m, ε_r = 3.5, diameter = 2mm) with a 2 mm exposed tip in ANSYS Electronic Desktop 2019 R1 HFSS (ANSYS, Canonsburg, PA). All lead models were 40 cm in length, consistent with the lengths of Abbott lead models 6172 and 6173. Each constructed model was inspected and adjusted if necessary, to ensure there were no self-intersections between the conductive wire and the insulation. Fig. 1 provides the steps of image segmentation and model creation.

Fig. 1.

Image segmentation and simplified DBS lead model construction. (a) and (b) 3D surface rendered view of CT images of a patient with implanted DBS leads. (c) Manual segmentation of lead trajectories. Preliminary 3D surfaces of lead trajectories were created (yellow), and points along lead trajectories (red) were extracted in 3D Slicer. (d) Alignment of 3D model of patient’s head (green) with the homogeneous MIDA model. (e) Lead trajectories were reconstructed in ANSYS HFSS as simplified lead models (conductive wire (red) embedded with a urethane insulating sheath (blue) with a 2 mm exposed tip). (f) Superposition of all 83 patient-derived and 177 artificial lead trajectories in the homogeneous MIDA model.

It is shown that there is substantial patient-to-patient variation in the extracranial trajectory of DBS leads, primarily reflecting the surgeon’s practice style [30]. To account for this variation and to include trajectories that are known to generate worst-case RF heating during DBS imaging [31], [32], we augmented the original dataset with 177 artificial DBS lead trajectories. Artificial DBS lead trajectories were constructed by manually drawing trajectory lines corresponding to the extracranial portion of the trajectory in Rhino 3D (Robert McNeal and Associates, Seattle, WA). These trajectory lines were drawn without references to patients’ postoperative CT images but followed general surgical practice as observed in our institution. The coordinates of points corresponding to the intracranial portions of artificial trajectories, that is, the segment from the lead’s tip to the exit point on the skull (~7–8 cm), were extracted from five actual DBS patients. Subsequently, the extracranial component was randomly connected to one of the five intracranial segments. Finally, trajectory lines for the leads were exported to ANSYS HFSS to reconstruct the conductive wire and insulation as previously described.

B. Electromagnetic Simulations

Electromagnetic simulations were implemented in ANSYS Electronic Desktop 2019 R1 HFSS. A numerical model of a 16rung low-pass birdcage head coil (diameter = 356 mm, length = 292 mm) tuned to the operating RF frequency of 64 MHz for 1.5 T MRI with quadrature excitation was loaded with the standard homogeneous head model and lead models for simulations. Additional simulations with patient-derived head models were performed to evaluate the effect of head variabilities on 1gSAR_max (Supplementary Fig. S1 and S2). A fixed voltage of 100 V was applied to each port of the coil in each simulation, generating a B₁⁺ of approximately 1.1 μT on a central transverse plane passing through the center of the head.

The 1gSAR_max was calculated in a 20 mm × 20 mm × 20 mm cubic region surrounding each lead-tip using the SAR calculation module incorporated in ANSYS HFSS per the IEEE STD P1 528.4 recommendation. These maximum SAR values represent the simulated ground-truth SAR values used to train the ML algorithm.

Simulations were performed with an adaptive mesh scheme. A fine resolution was applied to the leads with a maximum tetrahedral mesh edge length of 0.5 mm. Additionally, the 20 mm × 20 mm × 20 mm cubic region surrounding the lead-tip had a maximum tetrahedral mesh edge length of 2 mm. Iterative simulations were completed until the maximum difference in S-parameter values, ΔS, no longer exceeded the limit of 0.02. All simulations converged within 3 adaptive passes. Fig. 2 gives details of the simulation setup and mesh statistics.

Fig. 2.

Electromagnetic simulation setup. (a) Position of the MIDA head model and two example leads (blue) in a birdcage MRI head coil. 1gSAR_max was calculated in the 20 mm × 20 mm × 20 mm region surrounding the lead’s tip (green). (b) Example meshes of a lead (blue insulating sheath and red conductive wire) and the region surrounding the lead’s tip (green). (c) 1 g-averaged SAR displayed on an axial plane intersecting the lead-tips. (d) Details of mesh statistics for the example simulation in this figure.

To determine the incident electric field distribution along each lead trajectory, EM simulations were performed once with the head coil loaded with the homogeneous head model without the implanted leads. The tangential component of the incident electric field along each lead trajectory at a certain time point was then calculated and sampled at 5 mm increments along the length of each lead for a total of 80 E_tan values (Fig. 3). Each E_tan value served as an individual input feature for a total of 80 features used in the ML algorithm.

Fig. 3.

(a) Superposition of E_tan distributions along different lead trajectories in the homogeneous MIDA model. (b) Distribution of E_tan (color field) and the incident electric field (arrows) for lead trajectories that demonstrate low and high 1gSAR_max values, and (c) the evolution of E_tan along the length of the leads at t = 0.

C. Dataset Generation

Lead trajectories were randomly selected for training, validation, and testing; 80% of lead models were used for the training and validation data while the testing data was comprised of the remaining 20% of lead models. This procedure was repeated five times such that all data were used as test sets (called five-fold cross-validation).

D. Neural Network Details

The neural network architecture was implemented in Python 3.6 and Keras 2.3.1 with TensorFlow backend. Other Python libraries including Scikit-learn 0.21.2 were also used to support data processing [33]. We trained the algorithm to predict maximum SAR at the tips of implanted leads using the E_tan values along the leads. Predictive performance was evaluated with the correlation coefficient (R) and the root mean square error (RMSE) of the predicted SAR values from the ML approach compared to the ground-truth SAR values that were calculated from the full-wave EM simulations. Training loss per training epoch was collected and analyzed to assess potential overfitting to the data, as indicated by the MSE loss function

$$MSE=\frac{\sum _{i=1}^N {\left(y_i-{\widehat{y}}_i\right)}^2}{N}$$ (1)

where N (N = 208) is the number of samples, y_i is the SAR value from simulations of the i-th sample, and <di> is the predicted SAR value from the neural network. To further interrogate this possibility, we utilized a five-fold cross validation schema which allowed us to evaluate the model’s training and testing on a unique test dataset for five different models using our dataset. A preliminary hyperparameter optimization was performed to determine the optimal number of hidden layers, neurons per hidden layer, and epoch. Hyperparameter optimization was conducted using the GridSearchCV method from the Scikit-learn library with fivefold cross validation on the training and validation dataset of the first fold. During GridSearchCV, we evaluated neural network models with two to four hidden layers and the number of neurons per hidden layer between the number of inputs and outputs [20]. Optimal hyperparameter values were selected from the neural network model that minimized MSE while improving R during GridSearchCV. Once the optimal hyperparameter values were identified, we froze these values and applied them to the four remaining unique training and testing datasets in the five-fold cross validation schema.

III. Results

A. Simulated Maximum SAR at the Lead-tip

RF-induced power deposition in the tissue was quantified with 1gSAR_max, which occurred at the lead-tip for all models. 1gSAR_max ranged from 1.6 W/kg to 2674.6 W/kg across patient-derived and artificial lead trajectories, with leads with artificial trajectories demonstrating significantly higher 1gSAR_max (247.4 ± 388.9 W/kg) than leads with patient-derived trajectories (86.5 ± 65.3 W/kg). The highest 1gSAR_max for an artificial trajectory was 2674.6 W/kg compared to 283.4 W/kg for a patient-derived trajectory (Fig. 4). Thus, inclusion of artificial trajectories extended the range of SAR values in the dataset to illustrate worst-case scenario heating.

Fig. 4.

Distribution of maximum 1 g-averaged SAR at the lead-tip of 260 different lead trajectories from EM simulations.

B. Neural Network Architecture and Model Performance

In this study, we demonstrated that even a not-so-deep neural network architecture can be beneficial for predicting RF-induced power deposition in the tissue as indicated by SAR. The final model was a fully connected feed-forward neural network comprised of three hidden layers with 70, 70, and 60 nodes in the first, second, and third layer, respectively, and an output layer (Fig. 5). ReLU activation functions were used in the hidden layers while a linear activation function was used in the output layer. Network weights were determined with the Adam optimization routine [34]. Further, the effect of lead model sample size on model prediction performance was investigated. The following sample sizes were evaluated: 50, 100, and 150 lead models. For each dataset of varying sample sizes, the lead models were randomly partitioned into 80% of leads for training and validation and 20% for testing. Increasing the sample size from 50 lead models to 150 lead models led to improved predictions of 1gSAR_max during testing (RMSE = 357 W/kg to 239 W/kg) (Fig. 6). Further improvements were shown with the expanded dataset of 260 lead models.

Fig. 5.

Architecture of the final neural network. The fully connected neural network consisted of three hidden layers with 70, 70, and 60 hidden nodes in the first, second, and third hidden layer, respectively. E_tan values along the leads served as the inputs to the neural network with the output being the 1gSAR_max at the DBS lead-tip.

Fig. 6.

The effect of dataset size on the performance of the neural network for predicting 1gSAR_max during testing. Predicted 1gSAR_max values from testing the neural network plotted against the ground-truth 1gSAR_max values from EM simulations. The following sample sizes of (a) 50, (b) 100, and (c) 150 lead models were explored.

Training of the neural network was completed within minutes, and predictions were generated within seconds. After 250 epochs of training during each fold of five-fold cross validation, the ML algorithm performed with an average RMSE = 5.2 W/kg (Fig. 7). Fig. 8 demonstrates the performance of the neural network for predicting maximum SAR from the testing datasets based on the 80 E_tan values per lead trajectory. The ML algorithm was able to recognize high and low SAR producing trajectories during testing; testing performance is indicated with an average R = 0.82 and an average RMSE = 168 W/kg across all folds of the five-fold cross-validation schema.

Fig. 7.

Training history, indicated by the mean squared error loss function, for all folds of five-fold cross-validation. After 250 epochs of training during each fold, the neural network had an average RMSE = 5.2 W/kg.

Fig. 8.

Performance of the neural network with the test datasets for all folds of five-fold cross-validation. Predicted 1gSAR_max values from testing the neural network plotted against the ground-truth 1gSAR_max values from EM simulations.

IV. DISCUSSION

RF-induced heating in patients with active implants, such as DBS devices, is a significant safety concern given the increasing prevalence of patients with these implants and their need for MRI exams. Computational modeling using full-wave EM simulations are regulatory recommended methods for evaluating RF heating as they allow for characterization of the implant, patient model, and RF coil. However, covering the parameter space for clinically relevant configurations can require hundreds of thousands of simulations, and these simulations are notoriously intensive in terms of time and computational resources. This is currently a problematic barrier for developing novel MRI-friendly implants and imaging methodologies [35]–[41], and for stakeholders who strive to devise guidelines for safe imaging of existing devices [25]. Application of machine learning could potentially mitigate limitations of full-wave EM simulations by reducing the necessary computational resources.

In this proof-of-concept study, we explored the possibility of applying a ML-based prediction model to estimate power absorption in the tissue around tips of elongated implants (such as DBS leads), when only the background electric field of the MRI RF coil and the implant’s trajectory are available. We focused on simplified models of DBS leads as their MRI safety assessment is the most challenging due to the large range of observed RF heating values depending on the implant’s trajectory [24], [42]. This is because in contrast to the intracranial trajectory of leads, for which almost every neurosurgeon follows textbook guidelines to determine the lead trajectory, there are no rules or guidelines for placing the extracranial portion of the lead. As a result, surgeons can position the leads arbitrarily, leading to substantial patient-to-patient variation in the extracranial lead trajectories. For this reason, it is imperative to develop training datasets that include both patient-derived data, which are clinically relevant, as well as synthetic trajectories to account for inter-surgeon variability, outliers, and worst-case scenarios. Herein, we developed a dataset of 260 patient-derived and artificial lead models for computational modeling and simulations. Using these simulations, we subsequently trained a fully connected neural network to predict 1gSAR_max at the lead-tip from the respective distribution of E_tan along the length of the lead. The DBS lead trajectory is a known contributing factor to SAR amplification following MRI RF exposure [23], [25]. This factor is consistent with our neural network performance on the test dataset. The final neural network model recognized trajectories with either high or low 1gSAR_max. Due to the large range of SAR in the dataset and the magnitude of these values, some overestimation of SAR likely skewed the performance of the neural network, resulting in a high RMSE value.

While the presented neural network is capable of predicting maximum SAR at the tips of implanted leads given the distribution of E_tan along the lead trajectory, several limitations remain due to the proof-of-concept nature of our study. First, prediction of a single SAR value (i.e., the maximum SAR) is not sufficient to predict the temperature rise in the tissue. Future work should alternatively include thermal simulations to quantify temperature rise in the tissue and to train the network to directly predict the temperature which is the safety parameter of interest. Also, simulations were performed with simplified DBS lead models of a certain length and in a homogeneous head model. Additionally, only isolated DBS leads were evaluated in this study. Simulating the full DBS system—inclusive of the lead, extension, and implantable pulse generator (IPG)—could alter the range of SAR values. Whether the same neural network can be trained to predict SAR at the tips of leads with different lengths needs to be investigated in the future. This study was also limited to cases at 1.5 T MRI, and SAR amplification during MRI at higher fields (e.g., 3 T and above) where the electric and magnetic fields are less homogeneous remain to be explored. Furthermore, when more data becomes available, we will explore different, more modern neural network architectures to compare robustness, generalization, and fairness of the prediction models. The most challenging problems are handled by over-parametrized deep network architectures similar to ResNet, DenseNet, or an extension of such. We anticipate that since the problem of RF power absorption estimation was possible with a neural network model when the features were sampled from the lead trajectories, there is a tremendous opportunity to explore the same problem in a multi-center, multi-vendor setting.