Open Access
Add to BibSonomy
Abstract
The efficacy of catheter ablation treatment for atrial fibrillation is directly impacted by the quality of lesion formation. Two parameters that are critical for maximizing energy delivery are sustained catheter contact and orientation. Currently, these parameters must be inferred indirectly through tactile feedback or measurements of bioelectrical impedance and tip force. In this work, we propose a method for discerning contact and orientation based on direct endomyocardial imaging mediated by optical coherence tomography (OCT)-integrated ablation catheters. A two-stage classifier is developed to deduce contact parameters from M-mode images. Experimental validation within swine left-atrial specimens demonstrate accuracies of 99.96% and 92.88% for contact and orientation stages, respectively. These results highlight the potential of OCT M-mode imaging for guiding catheter placement during radiofrequency ablation interventions.
© 2019 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. INTRODUCTION
Percutaneous radiofrequency ablation (RFA) therapy is widely utilized for treatment of atrial fibrillation and is a promising alternative when pharmacological interventions have failed [1]. A fundamental goal of RFA treatment is the restoration of sinus rhythm through the electrical detachment of normal heart tissue from arrhythmogenic, pathological sites. Once ablation targets have been identified, catheter operators must establish sustained contact, under cardiac motion, throughout the duration of RF energy application. Current catheter tracking technologies detect the catheter tip position within a virtual, static anatomical shell. Although a stable reference devoid of motion aids in guiding catheter placement, it is often at the expense of registering the precise real-time tip electrode position relative to the beating heart [2]. Consequently, the quality of contact stability and orientation is often inferred from indirect parameters such as bioelectrical impedance, tactile feedback, or tip force [3,4]. Direct imaging from the catheter tip might help to better inform of interrupted contact and suboptimal orientation during RFA that could limit energy transfer and thus lesion quality. Furthermore, the likelihood of arrhythmia recurrence has been associated in part with anatomical region [5,6]. Such a device could assist in catheter micro-positioning in these problematic areas as well as within complex left-atrial geometries.
Optical coherence tomography (OCT) is an emerging high-resolution imaging modality that has been previously used for assessing cardiac structural substrates [7–9]. Additionally, catheter-based derivations of OCT have been applied for monitoring RF lesion delivery [10–13]. In this work, we sought to explore the use of OCT-integrated RF catheter imaging to inform on contact and orientation assessment. Utilizing a convolutional neural network (CNN)-based approach, we propose a technique for inferring the status of contact and orientation based on M-mode OCT images and validate it within swine left-atrial samples in vitro. The technique shows promise in distinguishing between contact and noncontact, in addition to tracking large differences within contact angle.
2. METHODS
A. System Setup
A grin-lens terminated single-mode fiber was inserted through the irrigation channel of a commercial RFA catheter (Thermocool, Biosense Webster, Diamond Bar, CA) to enable endomyocardial M-mode imaging through an opening drilled at the catheter tip. A reference arm was fabricated to match the integrated catheter sample arm, and both were integrated into a commercial spectral-domain OCT system centered at 1325 nm (Telesto I, Thorlabs Inc, Newton, NJ). The axial and lateral resolutions (prior to catheter mounting) were 7.5 μm and 18.5 μm in air, respectively. A 28 kHz line rate was used for M-mode acquisition with an averaging factor of 5. Data were acquired and processed using a custom program developed in C++ using the SpectralRadar SDK (Thorlabs, Newton, NJ). A schematic diagram of the system is shown in Fig. 1.
Fig. 1. Schematic diagram of the M-mode integrated radiofrequency ablation catheter. FC, fiber coupler; SMF, single-mode fiber; ODL, optical delay line; SLD, superluminescent diode; GL, grin lens; C, collimator; M, mirror.
B. Algorithm Overview and Dataset Acquisition
Algorithm development was based on OCT M-mode images acquired from the RF catheter tip. Each column corresponded to a sample reflectivity depth profile at a given time point. Units of 16 consecutive A-lines were used to compose a 2D image for input into the algorithm. This small batch size was selected in order to ensure rapid acquisition and subsequent prediction to minimize susceptibility to motion artifacts. The objective of the algorithm was to predict the status of catheter–tissue contact and orientation given a limited group of A-lines. This was accomplished using a two-stage classifier that consisted of two binary sub-classifiers. The first sub-classifier was for discriminating contact from non-contact scenarios. Following contact affirmation, the input image was directed to the second sub-classifier, which distinguished between two angle groups: large angle (greater than 30 deg) and small angle (less than or equal to 30 deg). Unless otherwise stated, all angles presented in this work were measured with respect to the surface parallel. A flow chart depicting the algorithm pipeline is shown in Fig. 2. Each sub-classifier was based on CNNs and implemented using the Keras [14] deep learning framework with TensorFlow [15] (Google LLC, Mountain View, CA) backend.
Fig. 2. Implementation of algorithm for guiding lesion delivery. (a) Envisioned integration of the algorithm within the RF ablation work flow. (b) Data pipeline for the proposed two-stage classification of catheter placement. Input batches of A-lines are first used to determine established contact. Once contact has been confirmed, the algorithm determines whether the catheter is oriented at a large or small angle to the tissue surface.
Left-atrial wedges were dissected from 10 healthy swine hearts and submerged in whole blood. All swine hearts were acquired fresh from a nearby butcher (Green Village Packing Co., Green Village, NJ) and delivered within 24 h of sacrifice. Data from the first three hearts were used for the contact sub-classifier stage, while the remaining hearts were utilized for the orientation sub-classification stage. Prior to this dataset acquisition, two separate hearts were acquired for system testing and protocol optimization. In generating the contact dataset, the catheter was oriented at three different angles with the tissue surface. For each angle, a micro-manipulator was used to adjust the angle and establish either contact or non-contact (blood interposed between tip and tissue surface) scenarios. In some cases, specimen motion was induced to mimic cardiac contraction. Several positions across each specimen were measured in order to account for the spatial variability in anatomical features within left-atrial samples. A total of 720,000 A-lines constituted the contact classifier dataset. Sample acquisitions are shown in Fig. 3.
Fig. 3. Example OCT M-mode images of contact scenarios. (a) Representative non-contact image when the catheter is about 5 mm or more away from the sample surface. (b) Example image when the catheter is less than 1 mm away from the sample surface. Though not in contact, the endocardium is easily visible through the blood layer. (c) When the catheter is in contact with the left atrium under translational sample motion. Bar height and width correspond to 100 μm and 5 ms, respectively.
Left-atrial samples from seven fresh swine hearts were used for the orientation sub-classifier; five were used to train the classifier, and two were used for testing. A similar experimental setup was used as the data for contact assessment. However, due to the undulated nature of the heart, data acquisition was limited to the relatively flat regions of the specimen as previously described [16]. Figure 4 describes the challenge introduced by variations in surface topology. Each atrial wedge sample was first flattened and secured to a corkboard submerged in a container of whole blood. The container was then fixed on an automated stage to acquire M-mode images at 100 different positions across a predefined flat region. For each position, 8,000 A-lines were acquired for each angle, and a total of three angles were acquired: 30 deg, 60 deg, and 90 deg from the tissue parallel. Images from 60 deg and 90 deg angles constituted the large angle orientation group, while 30 deg images made up the small angle group. Representative M-mode images are shown for each cohort in Fig. 5. This data acquisition scheme was utilized to enable future algorithm extension from binary to ternary orientation discrimination. In this work, however, because larger angles are preferred in practice, the dataset for large angular group was twice as large as the small angle group. A confusion matrix approach was used to evaluate the extent of bias due to unbalanced data sizes across classes [8,17]. Prior studies utilizing CNNs for breast cancer detection within an unbalanced set of histological images utilized such an approach to assess classification performance trained from disproportionate data [17].
Fig. 4. Ideal (a) and actual (b) sample catheter tissue orientations. Discrepancies induced by surface topology add uncertainty X in precise angle estimation. Image adapted from [16].
Fig. 5. Representative images from the large (a) and small (b) angle categories of catheter orientation used in this study. Bar corresponds to 5 ms.
The collected OCT M-mode images were partitioned into training, validation, and test sets. In order to fulfill the test speed requirement for real-time assessment, M-mode images were considered in batches of 16 A-lines. Thus, an image size of was used for the classifier input. Table 1 shows the allocation of images for model development for each classification stage. In the contact sub-classifier dataset, 30,000 images were used for training, while the remaining data were split equally to form the validation (7,500 images) and testing sets (7,500 images). For the orientation classifier dataset, each of the seven hearts contained 150,000 images obtained from 100 different spatial positions in three different angular orientations. A five-fold cross validation experiment was conducted using five out of the seven hearts as shown in Fig. 6. For each fold, four of these five hearts were used for training. Images from the last heart served as the validation set, where 75,000 images were used. This process was repeated five times so that data from each heart would serve as the validation dataset once. A model was trained for each fold that was selected as the one that performed the best on its respective validation set throughout training. Additionally, for each fold, data from the two remaining unseen hearts were used as a test set to evaluate the performance for each model generated during cross-validation. The average accuracy across folds was computed as the cross-validated accuracy. The final model for the orientation sub-classifier was selected as the one out of the five models that obtained the best accuracy on the testing set.
Fig. 6. Approach for five-fold cross validation of the orientation sub-classification model. Tr, training; V, validation; Te, testing. H1–H7 correspond to the seven total hearts used during the process.
Fig. 7. Contact classifier architecture. Original M-mode image was cropped and sampled before insertion into the contact classifier. The basic unit enclosed within the bracket was repeated three times with the same filter size, pooling size for the max pooling layer, and activation function. However, the filter dimensions for the convolutional layers were varied. The convolutional output was flattened and fed into a fully connected (FC) layer. The final output was activated using a sigmoid function to yield the final prediction.
Table 1. Number of Images Used for Sub-Classifier Development
C. Contact Versus Non-Contact Sub-Classification
A shallow convolutional network architecture was used for the contact assessment sub-classifier, as shown in Fig. 7. It contained three convolutional layers, interleaved with max pooling. The kernel size for each convolutional layer was , and the filter dimensions were 32, 32, and 64 for the first, second, and third convolutional layers, respectively. Horizontal and vertical strides were set to a default value of 1, and no zero padding was used. Afterward, the output was then flattened and fed into a single fully connected layer followed by a dropout [18] layer with dropout rate of 20%. Finally, a sigmoid activation function was used to predict the probability of the input being either contact or non-contact. A stochastic gradient descent optimizer was employed, with a learning rate of 0.01 and momentum of 0.9. Unless otherwise stated, the binary cross-entropy loss function was used for all optimizers within this study. A batch size of 64 images was used, and the network was trained over a total of 100 epochs. The model that produced the best accuracy with the validation set was updated and stored throughout training and selected as the final model.
D. Contact Orientation Sub-Classification
It is difficult to differentiate between the large angle and small angle from visual inspection. In addition, the features of the atrial specimen vary largely within sample locations in addition to varying across atrial specimens from different swine. To account for these factors, we explored the use of a shallow residual network, as depicted in Fig. 8 [19]. The input was first fed into a zero padding layer followed by a convolutional layer, batch normalization [20] layer, and a max pooling layer. Afterward, an identity block (Fig. 9) with a convolutional layer in the shortcut was used to match the dimensions in the main path followed by two identity blocks as described in Ref. [19]. Both blocks skipped three convolutional and batch normalization layer combinations. This output was then subjected to average pooling followed by layer flattening. Subsequently, a dropout layer with 50% dropout rate followed by a fully connected layer was used. Finally, a sigmoid activation function was utilized to estimate the likelihood of the originating angle group (large or small) of the input data. Following preliminary experimentation testing a range of optimizers and hyperparameters, the Adadelta optimizer was chosen with an initial learning rate of 0.01, which previously demonstrated reasonable performance [21]. A 32 image batch size was used and the network was trained for a total of 100 epochs. Similar to the contact sub-classifier, the model that produced the greatest accuracy with the validation set was stored throughout training and selected as the final model.
Fig. 9. Two types of identity blocks. The first type (a) was used when the input and output had the same dimensions. The identity block in (b) was used to match the case of incongruent dimensions. Figure adapted from [19].
All CNN models were originally implemented in Python. To test in real time, the models were converted to C++ by using the Frugally Deep application programming interface [22]. The real-time experimental setup was similar to that used when acquiring the training dataset.
3. RESULTS AND DISCUSSION
The test set performance and online model prediction times for both contact and orientation sub-classifiers are shown in Table 2. The contact sub-classifier achieved close to 100% test accuracy, while the orientation sub-classifier yielded a five-fold cross-validated accuracy at 92.88%; the final orientation model obtained 95.62% accuracy on the test set. The times taken for model predictions of a given input test image were 6.8 ms and 8.3 ms for contact and orientation sub-classifiers, respectively. A possible reason for the reduced accuracy in orientation predictions could perhaps be explained by the lack of specificity of appreciable features. For instance, for a given tissue site, the largest differences between small angle and large angle images are the penetration depth and signal-to-noise ratio. For a particular specimen, the greatest penetration depth and signal-to-noise ratio are observed when the catheter is effectively normal to the sample surface. However, at oblique incidence, these features decrease in magnitude. While this effect was observed to be fairly consistent within each sample for a particular site, it was difficult to distinguish this effect from image variations induced by native local tissue properties [23,24].
Table 2. Test Accuracies for Contact and Orientation Classification Stages, along with Corresponding Prediction Times
Contact force is an important parameter that can affect the contact angle given the soft and deformable nature of cardiac tissue. Our initial real-time experiments were conducted without controlling for the contact force. However, when we pilot tested on an atrial specimen, the standard deviation among different locations was considerable. We considered this may perhaps be caused by differences in contact force and conducted several experiments to verify this. In these experiments, classification results were tested when the catheter was placed in gentle contact and in firm contact with the atrial specimen. The accuracy results showed high consistency between catheter contact situations when the catheter was in gentle contact, producing 89.97% accuracy, which was close to the offline testing accuracy. To further illustrate the influence of contact force, we conducted another experiment in which the orientation classifier output was computed while transitioning from gentle contact to firm. In this experiment, the catheter was initially in light contact with the specimen surface at a small angle (less than 30 deg). Increasing force was gradually applied to the sample by advancing the catheter further along the initial orientation plane. The classification result transitioned from predicting a small angle into systematically predicting a large angle (Fig. 10). This can be attributed to the soft, deformable nature of cardiac tissue, which conforms to the catheter tip under moderate force, essentially producing a large contact angle scenario. In the future, we would like to assess the impact of contact force and incorporate it within our model to improve orientation estimation.
Fig. 10. Real-time orientation prediction while contact force was varied. Small angles correspond to class 1 and large angle to class 0. Classification result gradually transitions from small to large angle as increased contact force is applied.
Although strong performances within sub-classification accuracies were observed, there are several aspects that warrant further exploration. First, the further optimization of contact quality classification speeds could better enable its use in downstream tissue analysis, such as RF treatment monitoring. In this work, the training was conducted on a separate computer using a consumer-grade graphics processing unit (GTX 1080, NVIDIA, Santa Clara, CA) and developed within the python-based Keras environment. For real-time implementation, the model required conversion to C++ for compatibility with the SpectralRadar SDK and integration with the OCT system acquisition code. Following this conversion utilizing the Frugally Deep framework, the reported prediction speeds were based on the C++ converted model running on an Intel i7-4770 processor. GPU-based model predictions could potentially improve classification speeds but are currently unsupported using the Frugally Deep framework and may require an alternative approach for integration with the OCT system. Nonetheless, we observed adequate speeds within this work suitable for real-time feedback for operator guidance. Second, due to the wide spatial variability within the left atrium (e.g., endocardial thickness), the features among specimens have large differences, and within each individual sample, there is considerable variability among different positions. The dataset used for training and validation of the orientation sub-classifier was derived from five swine hearts and tested with two additional hearts. In certain angles, we accounted for location differences by taking 100 different positions for each sample. However, more samples may be needed to capture the within-sample variability and could help to improve the orientation classifier accuracy.
Another future goal for the orientation sub-classifier is its extension from the two major angle groups into more resolved angular orientations. However, this may be challenging within M-mode images, where feature contrasts within left atrial images are dominated more by anatomical variability as opposed to differences seen from angular orientation influenced by Fresnel reflection. Future work will be aimed at modifying catheter designs to incorporate other orientation-specific contrast mechanisms. These changes may include multi-position measurement, e.g., by 2D scanning of the sample arm beam or the integration of multiple M-mode sites within the catheter tip. However, further assessment of the importance of precise orientation on lesion size would be needed to justify the added complexity. It is possible that a coarse assessment of angle, such as the method presented in this work, is sufficient for guiding tip manipulation in a practical setting.
4. CONCLUSION
In summary, contact quality and contact orientation are important factors that influence lesion formation and subsequently affect the efficacy of RFA therapy. These parameters are challenging to determine and are often inferred from indirect measurements. This study demonstrates the potential of endomyocardial M-mode imaging using integrated catheters, along with CNN classifiers, to provide intraprocedural assessment of contact parameters. Evaluation of contact and orientation could be performed rapidly with high accuracy enabling the technology for real-time applications. Intraoperative feedback of these parameters could potentially help to optimize and maintain quality catheter placement to improve the reliability and reproducibilty of lesion sets.
Funding
National Institute of Health (NIH) (1DP2HL127776-01); National Science Foundation (NSF) (1454365).
Acknowledgment
The authors would like to thank Theresa Lye, Ziyi Huang, Dr. Yu Gan, and Dr. Yuye Ling for helpful discussions regarding experimental design and algorithm development. The authors would also like to thank Dr. Elizabeth Olsen and Ching Lin for permission to use the fiber splicing equipment. The authors declare that they have no conflict of interest.
REFERENCES
1. H. Calkins, G. Hindricks, R. Cappato, Y.-H. Kim, E. B. Saad, L. Aguinaga, J. G. Akar, V. Badhwar, J. Brugada, J. Camm, P.-S. Chen, S.-A. Chen, M. K. Chung, J. Cosedis Nielsen, A. B. Curtis, D. W. Davies, J. D. Day, A. d’Avila, N. M. S. Natasja de Groot, L. Di Biase, M. Duytschaever, J. R. Edgerton, K. A. Ellenbogen, P. T. Ellinor, S. Ernst, G. Fenelon, E. P. Gerstenfeld, D. E. Haines, M. Haissaguerre, R. H. Helm, E. Hylek, W. M. Jackman, J. Jalife, J. M. Kalman, J. Kautzner, H. Kottkamp, K. H. Kuck, K. Kumagai, R. Lee, T. Lewalter, B. D. Lindsay, L. Macle, M. Mansour, F. E. Marchlinski, G. F. Michaud, H. Nakagawa, A. Natale, S. Nattel, K. Okumura, D. Packer, E. Pokushalov, M. R. Reynolds, P. Sanders, M. Scanavacca, R. Schilling, C. Tondo, H.-M. Tsao, A. Verma, D. J. Wilber, and T. Yamane, “2017 hrs/ehra/ecas/aphrs/solaece expert consensus statement on catheter and surgical ablation of atrial fibrillation,” EP Europace 20, e1–e160 (2018). [CrossRef]
2. E. S. Gang, B. L. Nguyen, Y. Shachar, L. G. Farkas, L. Farkas, B. Marx, D. Johnson, M. C. Fishbein, C. Gaudio, and S. J. Kim, “Dynamically shaped magnetic fields: initial animal validation of a new remote electrophysiology catheter guidance and control system,” Circ. Arrhythmia Electrophysiol. 4, 770–777 (2011). [CrossRef]
3. F. Perna, E. K. Heist, S. B. Danik, C. D. Barrett, J. N. Ruskin, and M. Mansour, “Assessment of catheter tip contact force resulting in cardiac perforation in swine atria using force sensing technology,” Circ. Arrhythmia Electrophysiol. 4, 218–224 (2011). [CrossRef]
4. H. Cao, S. Tungjitkusolmun, Y. B. Choy, J.-Z. Tsai, V. R. Vorperian, and J. G. Webster, “Using electrical impedance to predict catheter-endocardial contact during RF cardiac ablation,” IEEE Trans. Biomed. Eng. 49, 247–253 (2002). [CrossRef]
5. K. Rajappan, P. M. Kistler, M. J. Earley, G. T. Thomas, M. Izquierdo, S. C. E. Sporton, and R. J. Schilling, “Acute and chronic pulmonary vein reconnection after atrial fibrillation ablation: a prospective characterization of anatomical sites,” Pacing Clinical Electrophysiol. 31, 1598–1605 (2008). [CrossRef]
6. Y. Sotomi, T. Kikkawa, K. Inque, K. Tanaka, Y. Toyoshima, T. Oka, N. Tanaka, Y. Nozato, Y. Orihara, K. Iwakura, Y. Sakata, and K. Fujii, “Regional difference of optimal contact force to prevent acute pulmonary vein reconnection during radiofrequency catheter ablation for atrial fibrillation,” J. Cardiovasc. Electrophysiol. 25, 941–947 (2014). [CrossRef]
7. X. Yao, Y. Gan, C. C. Marboe, and C. P. Hendon, “Myocardial imaging using ultrahigh-resolution spectral domain optical coherence tomography,” J. Biomed. Opt. 21, 61006 (2016). [CrossRef]
8. Y. Gan, D. Tsay, S. B. Amir, C. C. Marboe, and C. P. Hendon, “Automated classification of optical coherence tomography images of human atrial tissue,” J. Biomed. Opt. 21, 101407 (2016). [CrossRef]
9. C. P. Fleming, K. J. Quan, H. Wang, G. Amit, and A. M. Rollins, “In vitro characterization of cardiac radiofrequency ablation lesions using optical coherence tomography,” Opt. Express 18, 3079–3092 (2010). [CrossRef]
10. D. Herranz, J. Lloret, S. Jiménez-Valero, J. L. Rubio-Guivernau, and E. Margallo-Balbás, “Novel catheter enabling simultaneous radiofrequency ablation and optical coherence reflectometry,” Biomed. Opt. Express 6, 3268–3275 (2015). [CrossRef]
11. C. P. Fleming, H. Wang, K. J. Quan, and A. M. Rollins, “Real-time monitoring of cardiac radio-frequency ablation lesion formation using an optical coherence tomography forward-imaging catheter,” J. Biomed. Opt. 15, 030516 (2010). [CrossRef]
12. X. Zhao, X. Fu, C. Blumenthal, Y. T. Wang, M. W. Jenkins, C. Snyder, M. Arruda, and A. M. Rollins, “Integrated RFA/PSOCT catheter for real-time guidance of cardiac radio-frequency ablation,” Biomed. Opt. Express 9, 6400–6411 (2018). [CrossRef]
13. C. Fleming, N. Rosenthal, A. Rollins, and M. Arruda, “First in vivo real-time imaging of endocardial RF ablation by optical coherence tomography,” Innov. Cardiac Rhythm Manage. 201, 199 (2011).
14. F. Chollet, “keras,” 2015, https://github.com/fchollet/keras.
15. M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Isard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Mané, R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V. Vanhoucke, V. Vasudevan, F. Viégas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, and X. Zheng, “TensorFlow: large-scale machine learning on heterogeneous systems,” 2015, https://www.tensorflow.org/.
16. M. Zaryab, R. P. Singh-Moon, and C. P. Hendon, “Robust classification of contact orientation between tissue and an integrated spectroscopy and radiofrequency ablation catheter,” Proc.SPIE 10042, 100420O (2017). [CrossRef]
17. D. Bardou, K. Zhang, and S. M. Ahmad, “Classification of breast cancer based on histology images using convolutional neural networks,” IEEE Access 6, 24680–24693 (2018). [CrossRef]
18. N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: a simple way to prevent neural networks from overfitting,” J. Mach. Learn. Res. 15, 1929–1958 (2014).
19. K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” CoRR abs/1512.03385 (2015).
20. S. Ioffe and C. Szegedy, “Batch normalization: accelerating deep network training by reducing internal covariate shift,” in ICML (2015).
21. M. D. Zeiler, “ADADELTA: an adaptive learning rate method,” CoRR abs/1212.5701 (2012).
22. D. T. Hermann, “Frugally-deep,” 2018, https://github.com/Dobiasd/frugally-deep.
23. J. Pandian, D. Kaur, S. Yalagudri, S. Devidutta, G. Sundar, S. Chennapragada, and C. Narasimhan, “Safety and efficacy of epicardial approach to catheter ablation of ventricular tachycardia- an institutional experience,” Indian Heart J. 69, 170–175 (2017). [CrossRef]
24. A. D’Avila, P. Gutierrez, M. Scanavacca, V. Reddy, D. L. Lustgarten, E. Sosa, and J. A. F. Ramires, “Effects of radiofrequency pulses delivered in the vicinity of the coronary arteries: implications for nonsurgical transthoracic epicardial catheter ablation to treat ventricular tachycardia,” Pacing Clinical Electrophysio. 25, 1488–1495 (2002). [CrossRef]
