Open Access
Add to BibSonomy
Abstract
Radiofrequency ablation (RFA) is a minimally invasive procedure that is commonly used for the treatment of atrial fibrillation. However, it is associated with a significant risk of arrhythmia recurrence and complications owing to the lack of direct visualization of cardiac substrates and real-time feedback on ablation lesion transmurality. Within this manuscript, we present an automated deep learning framework for in vivo intracardiac optical coherence tomography (OCT) analysis of swine left atria. Our model can accurately identify cardiac substrates, monitor catheter-tissue contact stability, and assess lesion transmurality on both OCT intensity and polarization-sensitive OCT data. To the best of our knowledge, we have developed the first automatic framework for in vivo cardiac OCT analysis, which holds promise for real-time monitoring and guidance of cardiac RFA therapy..
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Atrial fibrillation is the most common sustained cardiac arrhythmia, which affects at least 2.3 million people in the United States [1–3]. With the aging population, the impact of atrial fibrillation will continue to rise with an estimation of 12–16 million people being affected by 2050 [4]. The prominent treatment of atrial fibrillation is through radiofrequency ablation (RFA), which could help to restore a regular heartbeat by creating non-conducting ablation lesions through the delivery of the thermal dose. The ablation lesions refer to tissue regions with irreversible injury that are damaged by radiofrequency energy. Single-lesion transmurality is essential to the success of the RFA treatment, as transmural lesions could disrupt the abnormal electrical pathways or isolate areas of abnormal electrical conduction. Herein, the knowledge of lesion transmurality could potentially optimize the treatment outcome [5]. Currently, RFA therapy is indirectly guided by monitoring temperature, impedance, and contact status through contact-force catheters or examining from low-resolution imaging modalities [6]. As a result, after an ablation procedure, atrial fibrillation recurrence remains greater than $20\%$ at 1-year-follow-up [5]. Therefore, for many patients, an additional procedure is required to achieve a chronic successful termination of atrial fibrillation.
Tissue type within the heart plays a critical role in the pathology of cardiovascular diseases [7,8]. With the knowledge of patients’ heart structure, the RFA treatment could potentially be optimized if real-time identification of ablation targets, assessment of catheter-tissue contact, and evaluation of acute lesion characteristics are available. Herein, cardiovascular imaging with direct assessment of cardiac substrates shows promise to improve the RFA procedure and lower the risk of atrial fibrillation recurrence and complications. Optical coherence tomography (OCT) is a depth-resolved imaging modality that can address these needs [9–14]. Previous research demonstrated that OCT could image the heart wall in real-time with a catheter-compatible fiber-optic probe via percutaneous access [15]. Polarization-sensitive OCT (PS-OCT) measures the endogenous contrast of birefringent tissue [16–18]. Integrated with an RFA catheter [19], PS-OCT is able to confirm lesion formation and monitor lesion transmurality in real-time during the RFA procedure [20,21].
Given the demand for real-time guidance on current RFA procedures, an automatic cardiac tissue analysis framework is in need to provide real-time image assessment of lesion quality and improve procedural outcomes. Over the past few years, deep learning has been extensively implemented within numerous clinical applications ranging from disease diagnosis to risk stratification [22–26]. Benefiting from the expansion of computational resources, deep learning is able to offer referral suggestions in real time. Recent advances have confirmed the potential of leveraging deep learning on OCT-based disease identification, prognosis prediction, and assessment of abnormal remodeling. In ophthalmology, deep learning has been widely adopted with superior performance on multiple clinical trials such as age-related macular degeneration [27], diabetic retinopathy [28,29], and glaucoma [30]. In addition to retinal disease analysis, previous studies also demonstrate the effectiveness of deep learning on breast cancer classification with OCT imaging [31]. These results lay the additional groundwork for using OCT for automatic tissue analysis through deep learning. There have been many examples of the use of deep learning within intravascular OCT, however applications of deep learning within the cardiac chambers are at the beginning stages. This manuscript demonstrates the usage of deep learning for cardiac ablation assessment with OCT.
In this study, we present an automatic framework for the analysis of in vivo intracardiac OCT images, aiming to improve RFA treatments. We demonstrate that our model can reliably classify cardiac tissue types, monitor catheter-tissue contact, and assess lesion transmurality within the atrium of in vivo swine models, showing promise for the guidance and monitoring of the atrial fibrillation ablation procedure with OCT. Our approach can work using OCT intensity/reflectivity images, and can be further improved by the incorporation of polarization-contrast, including birefringence and net retardance. The robustness of our approach is validated even in the presence of cardiac motion and unstable catheter-tissue contact. To the best of our knowledge, we developed the first deep learning model on in vivo cardiac OCT, which could enable automatic real-time assessment of catheter-tissue contact, cardiac tissue characterization, and lesion quality for direct RFA guidance.
2. Methods
Our study aims to improve RFA therapy for the treatment of atrial fibrillation by providing automated analysis of OCT images of the left atria to assess tissue types, contact, and lesion transmurality. As shown in Fig. 1, our approach consists of two major deep learning models, a segmentation network to generate pixel-wise cardiac substrates location masks and a lesion classification network to estimate the probability of lesion transmurality. The predicted tissue location masks can directly improve the treatment strategy by identifying the ablation targets and optimizing the treatment parameters through wall thickness measurement. During the ablation procedure, the catheter-tissue contact stability can also be monitored by calculating the proportion of pixels classified as blood in between the catheter tip and the cardiac wall over other cardiac tissues. The lesion classification network helps to identify the lesion transmurality, which could offer referral suggestions for the termination of the procedure. Our framework can incorporate multiple types of input data. Thus, in the evaluation section, we present the results from input w/o polarization-sensitive (PS) data: We denote V1 for models which only take OCT intensity images as input and V2 for models with OCT intensity images and PS data with the OCT Bscan stacked with the birefringence and net retardance.
Fig. 1. Algorithm flow of proposed framework. (A): Overall workflow of the proposed approach on ablation strategy improvement and ablation procedure monitoring; (B): Training process for the segmentation network; (C): Training process for the lesion classification network. Our framework consists of two deep learning models, a segmentation network and a lesion classification network. Both networks can incorporate multiple types of input data including OCT intensity images and PS data. The segmentation network could generate pixel-wise tissue location masks, which aim to provide information on ablation targets, wall thickness, and catheter-tissue contact status. The lesion classification network could monitor the lesion transmurality during the ablation to ensure a successful termination of the procedure.
2.1 Image acquisition
The dataset used within this study to develop and evaluate our framework was from an in vivo study of healthy pigs, previously reported in [32]. Briefly, the data was obtained in vivo from the left atrium of 4 healthy pigs weighing $66 \pm 4$ kg using an integrated PSOCT-RFA catheter. The imaging and the RFA treatment are colocally applied. During the in vivo experiments, we used continuous RFA treatment with clinical standard RFA parameters. The RFA treatment was conducted on an integrated PSOCT-RFA catheter and a standard radiofrequency generator (Maestro 3000, Boston Scientific, USA, temperature mode, power 40 W, temperature 65 $^{\circ } C$ ) with a range of ablation times (5, 15, 30, 45 s). The catheter was steered with an 8.5-Fr steering sheath (Agilis NxT Steerable introducer, St. Jude Medical, USA). A forward-imaging PS-OCT probe, which connected to the PS-OCT imaging system, was accommodated on the catheter to record the OCT data during the RFA treatment. The PS-OCT imaging system had a central wavelength of $1310 \ nm$, an axial resolution of $10 \ \mu m$, and a sensitivity of $100 \ dB$. The PS-OCT laser beam focuses along the longitudinal direction with lateral resolution around 17 $\mu m$. The tissue was imaged with a 0.7 $mm$ diameter scanning cone when the probe was rotated. The maximum penetration depth of PS-OCT in a living swine’s left atrium was about 1 $mm$. Raw images were acquired with 1500 A-lines per frame at a speed of 24 frames/s.
2.2 Cardiac tissue characterization
The segmentation network used in this study is modified from the classic biomedical segmentation network U-Net [33]. U-Net is an end-to-end image segmentation framework that achieves state-of-the-art performance in many biomedical segmentation tasks. To avoid overfitting, we adjusted the architecture and hyper-parameters according to the size of the dataset. Our segmentation model is formed by three encoder and decoder blocks with only one convolutional layer in each block. The number of channels in each convolutional layer is 64. The rectified linear activation function (ReLu) is a classic piecewise linear function that outputs the input directly if it is positive; otherwise, it outputs zero. The use of the RuLu function could help the network to learn the non-linear features from the model inputs. The max pooling layer could effectively downsample the feature maps generated from the convolutional layers, resulting in larger receptive fields and more robust local features with less translation invariance. To accelerate the training process, we use the batch normalization layer to standardize the layer inputs. We also add dropout layers in the inner encoder blocks for robust training. At the end of the classifier, we use the softmax activation function to normalize the output to a probability distribution over predicted tissue classes. A detailed description of our network architecture is shown in Fig. 1 (B).
Our segmentation network is jointly optimized by the combination of cross entropy loss (CEL) and Dice loss (DL) (Eqs. (1)–3):
The dataset used within the segmentation network included images with good catheter-tissue contact, unstable contact, and no contact from transmural and non-transmural lesions. A total of 205 images ($216 \times 1500$ pixels) were manually labeled with pixel-wise annotations into seven categories, including five tissue classes: endocardium, myocardium, adipose, collagenous, and blood, and two non-tissue classes: background and artifacts. The labeling process was blinded to the algorithm design. Due to the size of the dataset, we use the cross-validation strategy to evaluate our model. The samples were randomly divided into five cross-validation sets. During the training process, we cropped the OCT images and corresponding PS data into $216 \times 256$ overlapping patches to increase the size of the training set. The OCT images are normalized with horizontal flipping as the data augmentation.
Our segmentation network was trained from random initialization and optimized by the Adam optimizer [34] with an initial learning rate of 0.01. All experiments were conducted on a computer with the following features: an Intel core i9-9900K (16M Cache, up to 5.00 GHz) CPU and a RTX2080 GPU. During the cross-validation evaluation, all networks converged within 300 epochs.
2.3 Radiofrequency ablation lesion analysis
We develop a binary lesion classification network to provide real-time monitoring of lesion transmurality. The detection network consists of three basic blocks formed by the convolutional layer, batch normalization, and ReLu activation function. The number of channels in each convolutional layer is 16. A detailed description of the network architecture is shown in Fig. 1 (C). The lesion classification network is optimized by the cross entropy loss. During the lesion analysis phase, the probability of lesion transmurality is predicted from the classification network. The cross-A-line averaged birefringence was calculated by averaging the birefringence as a function of lateral distance within an individual frame.
The dataset for the radiofrequency ablation lesion classification algorithm training and evaluation consisted of 139 frames from 46 samples with good catheter-tissue contact. The frames were randomly sampled before and after the ablation procedure. Among 139 frames, 44 frames were validated as transmural lesions based on triphenyl tetrazolium chloride (TTC) staining [32]. The training/test sets were randomly split by frame to perform a five-fold cross validation. The detection network was formed by three basic blocks with the ReLu activation function and batch normalization layer to accelerate the training process. The networks were trained from random initialization and optimized by the Adam optimizer [34] with an initial learning rate of 0.01. All networks converged within 150 epochs. In the test stage, the threshold for lesion transmurality classification is 0.5.
2.4 Evaluation metrics
The performance of our multi-class segmentation network was evaluated using precision, recall, and F1 score (Eq. (4)).
For the radiofrequency lesion identification task, we used sensitivity, specificity, and accuracy to evaluate the lesion classification performance (Eq. (5)).
In addition to the evaluation metrics, we also used the ROC curve and corresponding AUC value to illustrate the overall lesion detection ability of our proposed classifier. The receiver operating characteristic (ROC) curve is a probability curve that plots the true positive rate (sensitivity) against the false positive rate (1-specificity) at various thresholds. The AUC value is calculated as the area under the ROC curve.
3. Results
3.1 Tissue segmentation
3.1.1 Substrate segmentation
To enable guidance, first a real-time method for substrate segmentation is needed. Precision, recall, and F1 score are used to quantitatively evaluate our segmentation model. Table 1 presents the cross validation evaluation results of our segmentation network, where V1 corresponds to the model developed using OCT intensity images only; and V2 corresponds to the model developed using OCT intensity images and corresponding PS channels. As shown, our model can accurately differentiate the cardiac tissue types from the in vivo OCT data, even though the data quality was impacted by motion and artifact. In particular, it achieves high performance (around $90 \%$) on myocardium and blood tissue identification, which demonstrates promise on wall thickness estimation and catheter-tissue contact stability assessment.
Table 1. Evaluation metrics (%) of tissue segmentation model.
3.1.2 Catheter-tissue contact
Automated analysis allows assessment of catheter-tissue contact. In Fig. 2, we show segmentation results from examples with different catheter-tissue contact statuses. As shown, our prediction results are highly consistent with the ground truth annotations, which further demonstrate the effectiveness of our approach on tissue characterization. In addition to the location of cardiac substrates, these prediction results can also indicate catheter-tissue contact stability and cardiac wall thickness. The catheter is not perpendicular to the tissue if the majority of the imaging regions belong to blood.
Fig. 2. Cardiac OCT image examples with different catheter-tissue contact status. (A) and (C): Samples with stable catheter-tissue contact; (B) and (D): Samples with unstable catheter-tissue contact. The catheter-tissue contact status can be directly indicated from the predicted segmentation mask. The scale bar is 200 $\mu m$.
3.1.3 Wall thickness
The cardiac wall thickness is critical for choosing appropriate ablation parameters. Wall thickness can be directly estimated by the number of pixels within an A-line classified as myocardium. Figure 3 shows scatter plots of the manually measured wall thickness versus calculated wall thickness, from images displaying stable catheter-tissue contact. As shown, the wall thickness is accurately predicted by both models with correlation coefficients up to 0.95. Figure 4 shows two representative results from cardiac samples with different wall thicknesses. In Fig. 4 (C) and (D), our models accurately track the wall thickness at different locations, even in the presence of motion and artifacts. These results validate the high performance of our approach on cardiac tissue analysis, showing promise for treatment strategy improvement.
Fig. 3. Scatter diagram for wall thickness measurement. The ground truth wall thickness vs. the predicted wall thickness from (A) V1 and (B) V2. Each point represents the average thickness of corresponding test images. For both models, the predicted wall thickness is highly consistent with the ground truth values. Shaded areas indicate 95% confidence intervals.
Fig. 4. Representative wall thickness results. (A) and (B): OCT intensity images overlapped with manual labels (blue) and predictions (red); (C) and (D): Cross-A-line thickness measurement on corresponding samples. Our proposed approach can successfully track the cardiac wall thickness at different locations. The scale bar is 200 $\mu m$.
3.2 Lesion detection and analysis
We demonstrate the robustness of our model by presenting representative results from different samples, including the transmural samples, non-transmural samples, and samples with unstable catheter-tissue contact in Figs. 5–7. Apart from the predicted lesion transmurality, we also show the values of impedance, cross-A-line averaged myocardial birefringence, and cross-A-line averaged myocardial intensity, as these parameters are commonly trusted for lesion transmurality evaluation. The cross-A-line data is obtained by averaging the entire B-scan through the x-axis and concatenating all averaged A-lines along the timeline, and thus it is a function of depth and time. Moreover, we also present the figure of birefringence and OCT Bscans before and after the RFA procedure to directly show the change of tissue regions over the procedure. The probability curves in Figs. 5–7 are obtained from V2. Figure 5 shows a representative transmural sample at the right pulmonary vein. Figure 5 (A) and (B) present the birefringence and OCT Bscan before the RFA procedure, while Fig. 5 (C) and (D) show the regions after the RFA procedure. Figure 5 (E) and (F) show the results of cross-A-line averaged birefringence and intensity values, while Fig. 5 (G) and (H) present cross-A-line averaged birefringence and intensity values within the myocardium regions. In comparison with Fig. 5 (E)/(F), Fig. 5 (G)/(H) highlights the birefringence/intensity values on the myocardium regions and shows the change of myocardial birefringence/intensity over the course of radiofrequency energy delivery. Figure 5 (I)/(J) averages the birefringence/intensity within the myocardium from Fig. 5 (G)/(H), resulting in a single (average) birefringence/intensity value for each frame. These figures provide a direct visualization of the trend of birefringence or intensity change with time and over the course of the RFA energy delivery. During the ablation, we can clearly observe a decrease of birefringence on myocardium regions in Fig. 5 (I), while the intensity values in Fig. 5 (J) are only slightly increased. In concordance with the observations in [32], the values of impedance have largely decreased at the beginning of the procedure in Fig. 5 (K). Figure 5 (L) presents the predicted probability of lesion transmurality. As shown, this prediction result indicates the change of lesion transmurality along the procedure, which can provide additional guidance on the ablation process.
Fig. 5. A representative example of the transmural lesions at the right pulmonary vein. (A) is the tissue birefringence before the RFA procedure (time 0s); (B) is the OCT Bscan before the RFA procedure (time 0s); (C) is the tissue birefringence after the RFA procedure (time 40s); (D) is the OCT Bscan after the RFA procedure (time 40s); The scale bar is 200 $\mu m$. (E) is the cross-A-line averaged tissue birefringence; (F) is the cross-A-line averaged OCT intensity; (G) is the cross-A-line averaged tissue birefringence on the myocardium region; (H) is the cross-A-line averaged OCT intensity on the myocardium region; (I) is the averaged tissue birefringence on the myocardium region over the course of RFA energy delivery; (J) is the averaged OCT intensity on the myocardium region over the course of RFA energy delivery; (K) is the impedance; (L) is the predicted probability of ablation obtained from V2. The decision threshold is marked by the yellow dash line.
Fig. 6. A representative example of non-transmural lesions at the right pulmonary vein. (A) is the tissue birefringence before the RFA procedure (time 0s); (B) is the OCT Bscan before the RFA procedure (time 0s); (C) is the tissue birefringence after the RFA procedure (time 30s); (D) is the OCT Bscan after the RFA procedure (time 30s); The scale bar is 200 $\mu m$. (E) is the cross-A-line averaged tissue birefringence; (F) is the cross-A-line averaged OCT intensity; (G) is the cross-A-line averaged tissue birefringence on the myocardium region; (H) is the cross-A-line averaged OCT intensity on the myocardium region; (I) is the averaged tissue birefringence on the myocardium region over the course of RFA energy delivery; (J) is the averaged OCT intensity on the myocardium region over the course of RFA energy delivery; (K) is the impedance; (L) is the predicted probability of ablation obtained from V2. The decision threshold is marked by the yellow dash line.
Fig. 7. A representative example with unstable catheter-tissue contact. (A) is the tissue birefringence before the RFA procedure (time 0s); (B) is the OCT Bscan before the RFA procedure (time 0s); (C) is the tissue birefringence after the RFA procedure (time 40s); (D) is the OCT Bscan after the RFA procedure (time 40s); The scale bar is 200 $\mu m$. (E) is the cross-A-line averaged tissue birefringence; (F) is the cross-A-line averaged OCT intensity; (G) is the cross-A-line averaged tissue birefringence on the myocardium region; (H) is the cross-A-line averaged OCT intensity on the myocardium region; (I) is the averaged tissue birefringence on the myocardium region over the course of RFA energy delivery; (J) is the averaged OCT intensity on the myocardium region over the course of RFA energy delivery; (K) is the impedance; (L) is the predicted probability of ablation obtained from V2. The decision threshold is marked by the yellow dash line.
In Fig. 6, we show the lesion analysis results from a non-transmural example. In comparison with Fig. 5, the changes of birefringence in Fig. 6 (I) are less obvious. In contrast, the impedance in Fig. 6 (K) has a similar changing pattern as in Fig. 5 (K). In Fig. 6 (L), the predicted lesion probability remains close to zero during the entire process, which accurately indicates the lesion transmurality status in the experiment. These results reveal the potential unreliability of using impedance on lesion identification and further demonstrate the effectiveness of applying OCT on RFA guidance.
Figure 7 presents the lesion analysis results from a sample with unstable catheter-tissue contact. In Fig. 7, all results appear in a fluctuation manner with large changes in both birefringence and OCT intensity. According to the TTC staining, this sample is identified as non-transmural, which is consistent with our prediction in Fig. 7 (L). These results evidently illustrate the robustness of our model in different situations.
Finally, we demonstrate the effectiveness of our model on lesion transmurality assessment by providing quantitative evaluations. Our approach is evaluated on the accuracy, sensitivity, and specificity via a five-fold cross-validation strategy. The samples are classified as transmural or non-transmural based on the OCT intensity images (V1) or OCT intensity images with PS data (V2). The ground truth labels were verified through TTC staining. Figure 8 shows the ROC curves, which describe the high performance of the proposed classifiers on lesion detection. In Fig. 8 (A) and (B), we show the ROC curves averaged from the five validation sets with shaded areas indicating standard deviation and red marks corresponding to the values reported in Table 2. In Fig. 8 (C) and (D), we report the detailed ROC curves from each validation set. In addition to ROC curves in Fig. 8, we provide the quantitative evaluation results of the proposed model on samples with stable catheter-tissue contact in Table 2. As shown, our model can accurately identify lesion transmurality with good performance on all evaluation metrics.
Fig. 8. ROC curves of two classifiers on lesion identification task. Classification results from (A) V1 (the proposed approach with intensity images as input) and (B) V2 (the proposed approach with intensity images and PS data as input). The red marks indicate the points on the ROC curves corresponding to the values reported in Table 2. (C) and (D) represent the ROC curves of each validation set obtained from model V1 and V2 correspondingly. Both classifiers achieve high performance on lesion transmurality classification. Shaded areas indicate standard deviation.
Table 2. Evaluation metrics (%) of lesion detection results.
4. Discussion
In this study, we present an in vivo cardiac analysis framework for RFA procedure improvement. Our approach consists of two deep learning models, a segmentation network and a lesion classification network. The segmentation network could generate pixel-wise tissue location masks to provide information of ablation targets, wall thickness, and catheter-tissue contact status. The lesion classification network could monitor the lesion transmurality to ensure a successful termination of the procedure. Validating on an in vivo OCT dataset, our model achieves high performance on both tissue identification (Precision > 80%) and lesion classification (AUC > 0.9). Moreover, our results also show that OCT could provide complementary information, when standard monitoring methods such as impedance are difficult to interpret.
Results from this study show the promise of automated analysis of left atrial OCT images to extract important features to improve RFA procedures. The information on the cardiac substrate can be used for the direct guidance of the RFA procedure. The predicted tissue location masks can directly improve the treatment strategy by identifying the ablation targets and optimizing the treatment parameters through wall thickness measurement. In the future, this information can be used to optimize the ablation parameters such as temperature, power, or energy delivery duration. During the ablation procedure, a perpendicular contact status is essential to the success of ablation, as it directly impacts the thermal dose delivered to the target regions. Without perpendicular catheter-tissue contact, the thermal dose received by the tissue regions might be insufficient to undergo irreversible injury, resulting in arrhythmia recurrences and repeat therapies. Hence, real-time catheter-tissue contact monitoring could help to optimize the therapy outcome during the RFA procedure. Based on the segmentation results, the catheter-tissue contact stability can be directly estimated by calculating the proportion of blood regions over other cardiac tissues in the OCT images. Thus, by monitoring tissue-blood proportion, the position of the catheter can be adjusted in time to ensure a stable energy delivery during the RFA treatment.
The lesion classification network helps to identify the lesion transmurality, which could offer referral suggestions for the termination of the procedure. In particular, the specificity reaches 95% for the lesion assessment module, indicating that few non-transmural cases are misclassified using our model. This performance holds promise to lower the chance of atrial fibrillation recurrence, as it helps to avoid non-transmural lesions during the treatment.
Current RFA treatment is guided by indirect parameters resulting in high atrial fibrillation recurrence. Our approach aims to address this challenge by providing real-time monitoring of cardiac wall structure, catheter-tissue contact status, and lesion transmurality assessment. In Fig. 2, we show the automatic segmentation results of multiple cardiac substrates with different catheter-tissue contact statuses. With tissue location information, the treatment strategy can be directly improved by optimizing ablation parameters and avoiding critical structures. From these results, we can also clearly observe the catheter-tissue contact stability, which is critical for the success of the ablation procedure, as unstable contact is one of the key reasons for treatment failure. Moreover, the ablation strategy can be further adjusted by the knowledge of cardiac wall thickness. In Fig. 5–7, we demonstrate the lesion analysis results from samples with different catheter-tissue contact statuses. In these examples, our model successfully demonstrates lesion transmurality during the RFA procedure. Overall, these results confirm the effectiveness of our proposed approach and show its robustness in the presence of cardiac motion, blood, and artifacts due to catheter-based imaging.
We will extend our work in the following aspects in the future. First, we will develop supporting software to enable real-time model implementation during the RFA procedure. In addition, we will also explore the potential of using deep learning models on the raw OCT data, which will further save the time of interfering with the raw OCT data into the OCT frames. Second, our current experiments are conducted on data obtained from healthy pigs. To further evaluate our framework, we will perform in vivo analysis on animal models with cardiovascular diseases such as atrial fibrillation. As been observed in [13], cardiac disease might alter the features of cardiac substrates, leading to heterogeneous heart remodeling and complex tissue types such as fibrofatty infiltration and fibrotic myocardium. In the future, we will update our model with the enriched dataset.
5. Conclusion
In this study, we developed the first deep learning based automatic lesion analysis framework for in vivo intracardiac OCT images of the left atria. Our approach provides automated segmentation and classification of cardiac tissue types, assessment of lesion transmurality, and catheter-tissue contact. In addition to the segmentation model, we develop a binary classification network to assess the lesion transmurality during the RFA procedure. By evaluating our methods using an in vivo cardiac OCT dataset, we demonstrate the effectiveness of our approach on lesion analysis and further indicate its potential for real-time RFA guidance.
Funding
Cheung-Kong Innovation Doctoral Fellowship; National Heart, Lung, and Blood Institute (NIH-5R01HL14936).
Disclosures
The authors declare no conflicts of interest.
Data Availability
The data that support the findings of this study are available from the corresponding author upon reasonable request.
References
1. Y.-K. Iwasaki, K. Nishida, T. Kato, and S. Nattel, “Atrial fibrillation pathophysiology: implications for management,” Circulation 124(20), 2264–2274 (2011). [CrossRef]
2. C. R. Wyndham, “Atrial fibrillation: the most common arrhythmia,” Texas Heart Inst. J. 27, 257 (2000).
3. S. Khurshid, S. H. Choi, L.-C. Weng, E. Y. Wang, L. Trinquart, E. J. Benjamin, P. T. Ellinor, and S. A. Lubitz, “Frequency of cardiac rhythm abnormalities in a half million adults,” Circ Arrhythm Electrophysiol. 11(7), e006273 (2018). [CrossRef]
4. Y. Miyasaka, M. E. Barnes, B. J. Gersh, S. S. Cha, K. R. Bailey, W. P. Abhayaratna, J. B. Seward, and T. S. Tsang, “Secular trends in incidence of atrial fibrillation in Olmsted county, Minnesota, 1980 to 2000, and implications on the projections for future prevalence,” Circulation 114(2), 119–125 (2006). [CrossRef]
5. C. A. Linte, J. J. Camp, M. E. Rettmann, D. Haemmerich, M. K. Aktas, D. T. Huang, D. L. Packer, and D. R. Holmes III, “Lesion modeling, characterization, and visualization for image-guided cardiac ablation therapy monitoring,” J. Med. Imaging 5(02), 1–021218 (2018). [CrossRef]
6. J. Sra, G. Narayan, D. Krum, A. Malloy, R. Cooley, A. Bhatia, A. Dhala, Z. Blanck, V. Nangia, and M. Akhtar, “Computed tomography-fluoroscopy image integration-guided catheter ablation of atrial fibrillation,” J. Cardiovasc. Electrophysiol. 18(4), 409–414 (2007). [CrossRef]
7. O. V. Aslanidi, M. R. Boyett, H. Dobrzynski, J. Li, and H. Zhang, “Mechanisms of transition from normal to reentrant electrical activity in a model of rabbit atrial tissue: interaction of tissue heterogeneity and anisotropy,” Biophys. J. 96(3), 798–817 (2009). [CrossRef]
8. M. Haissaguerre, A. J. Shah, H. Cochet, M. Hocini, R. Dubois, I. Efimov, E. Vigmond, O. Bernus, and N. Trayanova, “Intermittent drivers anchoring to structural heterogeneities as a major pathophysiological mechanism of human persistent atrial fibrillation,” The J. Physiol. 594(9), 2387–2398 (2016). [CrossRef]
9. 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(06), 1 (2016). [CrossRef]
10. C. J. Goergen, H. Radhakrishnan, S. Sakadžić, E. T. Mandeville, E. H. Lo, D. E. Sosnovik, and V. J. Srinivasan, “Optical coherence tractography using intrinsic contrast,” Opt. Lett. 37(18), 3882–3884 (2012). [CrossRef]
11. M. Gupta, A. M. Rollins, J. A. Izatt, and I. R. Efimov, “Imaging of the atrioventricular node using optical coherence tomography,” J. Cardiovasc. Electrophysiol. 13(1), 95 (2002). [CrossRef]
12. T. H. Lye, V. Iyer, C. C. Marboe, and C. P. Hendon, “Mapping the human pulmonary venoatrial junction with optical coherence tomography,” Biomed. Opt. Express 10(2), 434–448 (2019). [CrossRef]
13. Y. Gan, T. H. Lye, C. C. Marboe, and C. P. Hendon, “Characterization of the human myocardium by optical coherence tomography,” J. Biophotonics 12(12), e201900094 (2019). [CrossRef]
14. 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(3), 3079–3092 (2010). [CrossRef]
15. H. Wang, W. Kang, A. P. Bishop, A. M. Rollins, T. Carrigan, N. Rosenthal, and M. Arruda, “In vivo intracardiac optical coherence tomography imaging through percutaneous access: toward image-guided radio-frequency ablation,” J. Biomed. Opt. 16(11), 110505 (2011). [CrossRef]
16. M. R. Hee, D. Huang, E. A. Swanson, and J. G. Fujimoto, “Polarization-sensitive low-coherence reflectometer for birefringence characterization and ranging,” J. Opt. Soc. Am. B 9(6), 903–908 (1992). [CrossRef]
17. J. F. De Boer, C. K. Hitzenberger, and Y. Yasuno, “Polarization sensitive optical coherence tomography–a review,” Biomed. Opt. Express 8(3), 1838–1873 (2017). [CrossRef]
18. B. Baumann, “Polarization sensitive optical coherence tomography: a review of technology and applications,” Appl. Sci. 7(5), 474 (2017). [CrossRef]
19. 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(12), 6400–6411 (2018). [CrossRef]
20. D. Herranz, J. Lloret, S. Jiménez-Valero, J. Rubio-Guivernau, and E. Margallo-Balbás, “Novel catheter enabling simultaneous radiofrequency ablation and optical coherence reflectometry,” Biomed. Opt. Express 6(9), 3268–3275 (2015). [CrossRef]
21. X. Zhao, O. Kilinc, C. J. Blumenthal, D. Dosluoglu, M. W. Jenkins, C. S. Snyder, M. Arruda, and A. M. Rollins, “Intracardiac radiofrequency ablation in living swine guided by polarization-sensitive optical coherence tomography,” J. Biomed. Opt. 25(5), 056001 (2020). [CrossRef]
22. T. Davenport and R. Kalakota, “The potential for artificial intelligence in healthcare,” Future Healthcare J. 6(2), 94–98 (2019). [CrossRef]
23. Y. Peng, T. D. Keenan, Q. Chen, E. Agrón, A. Allot, W. T. Wong, E. Y. Chew, and Z. Lu, “Predicting risk of late age-related macular degeneration using deep learning,” npj Digit. Med. 3(1), 111 (2020). [CrossRef]
24. Y. Sun, J. Wang, J. Shi, and S. A. Boppart, “Synthetic polarization-sensitive optical coherence tomography by deep learning,” npj Digit. Med. 4(1), 105–107 (2021). [CrossRef]
25. Y. Deng, L. Lu, L. Aponte, A. M. Angelidi, V. Novak, G. E. Karniadakis, and C. S. Mantzoros, “Deep transfer learning and data augmentation improve glucose levels prediction in type 2 diabetes patients,” npj Digit. Med. 4(1), 109–113 (2021). [CrossRef]
26. A. Esteva, K. Chou, S. Yeung, N. Naik, A. Madani, A. Mottaghi, Y. Liu, E. Topol, J. Dean, and R. Socher, “Deep learning-enabled medical computer vision,” npj Digit. Med. 4(1), 5–9 (2021). [CrossRef]
27. D. Ferrara, R. E. Silver, R. N. Louzada, E. A. Novais, G. K. Collins, and J. M. Seddon, “Optical coherence tomography features preceding the onset of advanced age-related macular degeneration,” Invest. Ophthalmol. Vis. Sci. 58(9), 3519–3529 (2017). [CrossRef]
28. G. E. Lang, “Optical coherence tomography findings in diabetic retinopathy,” Diabetic Retinopathy 39, 31–47 (2007). [CrossRef]
29. J. Lee and R. Rosen, “Optical coherence tomography angiography in diabetes,” Curr. Diabetes Rep. 16(12), 123–127 (2016). [CrossRef]
30. I. I. Bussel, G. Wollstein, and J. S. Schuman, “OCT for glaucoma diagnosis, screening and detection of glaucoma progression,” Br. J. Ophthalmol. 98(Suppl 2), ii15–ii19 (2014). [CrossRef]
31. F. T. Nguyen, A. M. Zysk, E. J. Chaney, S. G. Adie, J. G. Kotynek, U. J. Oliphant, F. J. Bellafiore, K. M. Rowland, P. A. Johnson, and S. A. Boppart, “Optical coherence tomography: the intraoperative assessment of lymph nodes in breast cancer,” IEEE Eng. Med. Biol. Mag. 29(2), 63–70 (2010). [CrossRef]
32. X. Zhao, O. Ziv, R. Mohammadpour, B. Crosby, W. J. Hoyt, M. W. Jenkins, C. Snyder, C. Hendon, K. R. Laurita, and A. M. Rollins, “Polarization-sensitive optical coherence tomography monitoring of percutaneous radiofrequency ablation in left atrium of living swine,” Sci. Rep. 11(1), 24330 (2021). [CrossRef]
33. O. Ronneberger, P. Fischer, and T. Brox, “U-net: convolutional networks for biomedical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention (Springer, 2015), pp. 234–241.
34. D. P. Kingma and J. Ba, “Adam: maethod for stochastic optimization,” arXiv, arXiv:1412.6980 (2014). [CrossRef]
