Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Enabling Automated Device Size Selection for Transcatheter Aortic Valve Implantation

Enabling Automated Device Size Selection for Transcatheter Aortic Valve Implantation Hindawi Journal of Interventional Cardiology Volume 2019, Article ID 3591314, 7 pages https://doi.org/10.1155/2019/3591314 Research Article Enabling Automated Device Size Selection for Transcatheter Aortic Valve Implantation 1,2 1 3 4 5 Patricio Astudillo , Peter Mortier, Johan Bosmans, Ole De Backer, Peter de Jaegere, 1 2 Matthieu De Beule, and Joni Dambre FEops, Technologiepark-Zwijnaarde 122, Ghent, Belgium Department of Electronics and Information Systems, UGent-imec, Technologiepark-Zwijnaarde 126, Ghent, Belgium University Hospital Antwerp (UZA), Antwerp, Belgium Department of Cardiology, Rigshospitalet University Hospital, Copenhagen, Denmark Department of Cardiology, Erasmus MC, Rotterdam, Netherlands Correspondence should be addressed to Patricio Astudillo; patricio.astudillo@feops.com Received 29 April 2019; Revised 5 August 2019; Accepted 16 August 2019; Published 3 November 2019 Academic Editor: Amit Segev Copyright © 2019 Patricio Astudillo et al. �is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. �e number of transcatheter aortic valve implantation (TAVI) procedures is expected to increase signi‡cantly in the coming years. Improving eˆciency will become essential for experienced operators performing large TAVI volumes, while new operators will require training and may bene‡t from accurate support. In this work, we present a fast deep learning method that can predict aortic annulus perimeter and area automatically from aortic annular plane images. We propose a method combining two deep convolutional neural networks followed by a postprocessing step. �e models were trained with 355 patients using modern deep learning techniques, and the method was evaluated on another 118 patients. �e method was validated against an interoperator variability study of the same 118 patients. �e di‘erences between the manually obtained aortic annulus measurements and the automatic predictions were similar to the di‘erences between two independent observers (paired di‘. of 3.3 ± 16.8 mm vs. 1.3 ± 21.1 mm for the area and a paired di‘. of 0.6 ± 1.7 mm vs. 0.2 ± 2.5 mm for the perimeter). �e area and perimeter were used to retrieve the suggested prosthesis sizes for the Edwards Sapien 3 and the Medtronic Evolut device retrospectively. �e au- tomatically obtained device size selections accorded well with the device sizes selected by operator 1. �e total analysis time from aortic annular plane to prosthesis size was below one second. �is study showed that automated TAVI device size selection using the proposed method is fast, accurate, and reproducible. Comparison with the interobserver variability has shown the reliability of the strategy, and embedding this tool based on deep learning in the preoperative planning routine has the potential to increase the eˆciency while ensuring accuracy. �e number of TAVI procedures is increasing each year 1. Introduction rapidly [6], and considering the recent clinical data for low- Transcatheter aortic valve implantation (TAVI) has become risk patients will lead to an accelerated expansion in the the preferred treatment for patients with aortic stenosis at coming years. As a result, scalability of the complete pro- high risk for surgical aortic valve replacement (SAVR) [1]. cedure, including preoperative planning, becomes an im- Recently, it was concluded that, for intermediate-risk pa- portant aspect. Experienced operators can enlarge their tients, TAVI was similar to SAVR with respect to the pri- volume of TAVI cases, for example, by increasing pro- mary end-point of death or disabling stroke [2, 3]. Very cedural eˆciency. On the other hand, many new operators recent clinical data even show that TAVI is at least as good as will need to be trained, which logically leads to increased SAVR in low-risk patients [4, 5]. risks due to their limited experience. When focusing on the 2 Journal of Interventional Cardiology teach the neural network how to segment the aortic annular preoperative planning, accurate automated detection of the aortic annulus dimensions directly from multidetector plane using the ground truth annotations of the aortic annulus (Figure 1). computed tomography (MDCT) images could not only increase efficiency but also at the same time reduce operator &e deep learning model requires an architecture in variability, thereby minimizing the impact of experience on order to process the resampled and clipped aortic annular TAVI sizing. planes and compare the output of the model with the In this work, we present a deep learning method that can binary masks. &e used architecture was inspired by U-Net predict the aortic annulus perimeter and area automatically. [8] and deep residual nets [9] and consisted of two paths: a &e method is validated against an interoperator variability downscaling and an upscaling path. &e downscaling path study to assess its accuracy. As a final step, the impact of the extracted information from the aortic annular plane, and proposed method on the prosthesis size selection for both the upscaling path translated this information into a seg- mented aortic annulus. &e final sigmoid activation the Edwards Lifesciences and Medtronic transcatheter aortic bioprostheses was evaluated. function ensured that the output of the model contained probability values. &e details of the deep learning archi- tecture, training, and data-augmentation techniques are 2.Materials and Methods given in appendix A in Supplementary Materials (available 2.1. MDCT Imaging. &is retrospective study used the here). anonymized data of 473 patients collected from multiple &e deep learning architecture was used during the centra. &e mean age of this cohort was 80.82± 7.18 years, training phase to teach a deep learning model to segment the and 55% of the patients were female. &ere were 36 bicuspid aortic annulus from the aortic annular plane. patients in this cohort. &e patient data consisted of volu- metric MDCT images which were acquired to plan a TAVI procedure. &erefore, all MDCT images were contrast-en- 2.3.1. Training. Two models were trained using the training hanced and contained a certain degree of aortic stenosis. &e dataset and validated with the validation dataset. One model average row, column, and slice thickness of the MDCT was trained for each of the two resolutions (1 mm and images were 512.05 mm, 511.85 mm, and 0.83 mm. &e 0.5 mm) of the aortic annular planes. &e validation dataset aortic annular planes (AAP) were manually identified from consisted of the same 118 patients that were used for the the volumetric MDCT images using the standard method [7] interobserver variability study, and the training dataset and were used as input for this study. For this retrospective consisted of the remaining 355 patients. &e 36 bicuspid study, formal consent is not required. patients were distributed equally over the training and validation datasets. 2.2.ManualDetection. &e border of the aortic annulus was manually identified from the aortic annular planes by ob- 2.4.Detection. After training one model for each resolution, server 1. &e data of observer 1 were considered the ground a detection strategy was used to combine the output of both truth in this study. Observer 2 repeated this for 118 ran- models and to derive patient-specific anatomical in- domly selected patients in order to assess the interoperator formation: the area and perimeter of the aortic annulus. variability. Both observers applied the same manual method, &e detection of the area and perimeter of the aortic which consists of visual detection of the aortic annulus annulus of a single patient was performed in two steps: a within the AAP and annotating it using Mimics Innovation deep learning step and a postprocessing step. During the Suite 18 (Materialise, Leuven, Belgium). deep learning step, the aortic annular planes were ana- lysed by both models, and the output was combined and 2.3.AutomaticDetection. &is study aims at automating the normalized to a probability output that identified the manual segmentation and derives clinical patient-specific region of interest. During the postprocessing step, the contour of the region of interest was located with canny information as a postprocessing step. Preprocessing of the ground truth images and aortic annulus annotations were edge detection [10] from the probability output. &e area and perimeter were derived from this contour and serve necessary in order to prepare the data for training the deep learning models. as the final predicted output of the detection phase &e aortic annular planes were clipped and resampled in (Figure 2). order to fit the neural networks’ input. &e aortic annular As a final step, the derived aortic annulus dimensions were planes were resampled to an isotropic 1 mm resolution. As used to assess correct prosthesis size. &e perimeter was used to the deep learning network expected a 128 ×128 pixel plane select the proper Medtronic Evolut TAVR prosthesis size as input, the resampled aortic annular planes were clipped (https://www.medtronic.com/content/dam/medtronic-com/ around the center of the aortic annulus. A second isotropic products/cardiovascular/heart-valves-transcatheter/corevalve- 0.5 mm resolution was generated and clipped in the same evolut-r/documents/201709637EN-Evolut-PRO-TAV-in- SAV-Interactive-Sizing-Guide-FINAL.pdf), and the area was manner in order to double the level of segmentation detail. Cubic spline interpolation was used in order to retain the used to identify the Edwards Sapien 3 prosthesis size (https:// www.accessdata.fda.gov/cdrh_docs/pdf14/P140031c.pdf) sim- original Hounsfield units in the resampled aortic annular planes (Figure 1). Binary masks were generated in order to ilar to the manufacturer’s sizing matrix used in clinical practice. Journal of Interventional Cardiology 3 (a) (b) (c) (d) Figure 1: Examples of the aortic annular plane and the accompanying binary masks. &e resampled and clipped aortic annular planes (a) and the binary masks (b) with different resolutions, 1.0 mm (c) and 0.5 mm (d). 2.5.StatisticalAnalysis. &e Shapiro–Wilk test was performed patients for validation and observer variability assessment, it to test for normal distribution, and none of the predicted was possible to compare the method with both observers. distributions were normally distributed. Pearson correlation &e detection phase consisted of a deep learning phase coefficient was computed to evaluate the correlation between and a postprocessing phase. &e deep learning phase was model and both observers (with excellent correlation R > 0.9). validated by comparing the predicted segmentation (model) &e agreement between manual and the automatic landmark with the segmentation of both observers using the dice co- locations were evaluated using the nonparametric signed efficient. &e mean Dice score between model and observer 1 Wilcoxon test (with a significant p value <0.001). Bland– was 96% whereas the mean Dice score between both model Altman analysis for area and perimeter between model and and observer 2 and observer 1 and 2 was 89%. &e higher observer 1 and between both observers was performed. mean Dice score between model and observer 1 is expected because the model was trained with the data from observer 1. &e postprocessing phase derived the area and perimeter 2.6. Implementation. All the computational work was per- from the predicted segmentation and was validated by formed on a multicore computer with Titan X and P6000 comparing the predicted area and perimeter with the area GPUs (NVIDIA Corporation, Los Alamitos, CA). &e and perimeter of both observers. When comparing the models and the deep learning pipeline were developed with predicted anatomical measurements of the model with the PyTorch v0.4.1 [11]. data of both observers, there was no significant difference between the model and both observers for the area mea- 3.Results and Discussion surements. &e mean paired difference for all measurements 3.1. Results. &e proposed method trained two models, and was around zero, which means that the predicted anatomical the detection phase was validated using the 118 patients used measurements could be used in the same manner as the in the interoperator variability study. By using the same output of observer 1 or 2 (Table 1). 4 Journal of Interventional Cardiology Predict Compare Figure 2: A general overview of the method: the model predicts the probability plane from the original aortic annular plane. &e contours are detected, and the predicted area and perimeter are compared with the ground truth (GT). 2 Perimeter (mm) Area (mm ) 60 70 80 90 100 110 300 400 500 600 700 Observer 1 Observer 1 2 Model (R = 0.97) Model (R = 0.98) 2 Observer 2 (R = 0.94) Observer 2 (R = 0.97) (a) (b) Figure 3: Scatter plots comparing the interobserver correlation for the area (a) and perimeter (b). Excellent correction values were obtained between method’s ability to predict the correct prosthesis size model and observer 1 for the area (0.98) and perimeter (compared to both observers). &e predicted area and pe- (0.97). &e correlation values between observer 1 and 2 for rimeter were used to retrieve the Edwards Sapien 3 and the area (0.97) and perimeter (0.94) indicate that the manual Medtronic Evolut TAVR prosthesis sizes. &e automatically method is accurate (Figure 3). selected valve sizes were compared with valve sizes resulting Bland–Altman plots of the predicted and measured (ob- from the annular measurements of both observers. &e ratio server 1) area and perimeter are depicted in Figures 4 and 5. of agreement for Edwards Sapien 3 between model and both It is worth noting, when interpreting the Bland–Altman plots, observers is almost equal: 0.86 between model and observer that the model was repeatable since consecutive predictions 1 and 0.88 between both observers. &e ratio of agreement per patient yielded the same output. for the Medtronic Evolut TAVR prosthesis sizes between &e validation of the segmentation abilities and the area model and both observers is similar: 0.89 between model and and perimeter assessment were required to validate the observer 1 and 0.86 between both observers (Figure 6). Postprocess Model/observer 2 Model/observer 2 Journal of Interventional Cardiology 5 Table 1: A comparison of the anatomical measurements between model and both observers. Model vs. observer 1 Model vs. observer 2 Observer 1 vs. observer 2 Paired diff. p value Paired diff. p value Paired diff. p value 2) Area (mm 3.3± 16.8 0.008 2.0± 22.4 0.046 1.3± 21.1 0.752 Perimeter (cm) 0.6± 1.7 0.0001 0.5± 2.6 0.0016 0.2± 2.5 0.513 Paired difference reported as mean± standard deviation. 2 2 Area (mm ) Area (mm ) 100 100 50 50 0 0 –50 –50 –100 –100 300 400 500 600 700 300 400 500 600 700 Average of model and observer 1 Average of observer 2 and observer 1 (a) (b) Figure 4: Bland–Altman plots comparing the aortic annulus area for model vs. observer 1 (a) and both observers (b). Perimeter (mm) Perimeter (mm) 10 10 5 5 0 0 –5 –5 –10 –10 60 65 70 75 80 85 90 95 100 60 65 70 75 80 85 90 95 100 Average of model and observer 1 Average of observer 2 and observer 1 (a) (b) Figure 5: Bland–Altman plots comparing the aortic annulus perimeter for model vs. observer 1 (a) and both observers (b). Finally, it is relevant to report the processing time of the been validated on 118 patients to evaluate its accuracy, and manual and automated methods. &e automatic processing the results show that the area and perimeter can be predicted time from aortic annular plane to segmentation, anatomical in an automatic, reproducible, fast, and accurate way by measurement, and prosthesis size is below 1 second. combining the results of two networks followed by a postprocessing step. &e differences between the manually obtained aortic annulus measurements and the automatic 3.2. Discussion. In this work, an automated method is predictions are similar to the differences between two in- dependent observers, which indicates a satisfying accuracy proposed to facilitate and optimize the preoperative TAVI planning. It automatically predicts the area and perimeter of of the proposed approach. &e area and perimeter have also been used to retrieve the suggested prosthesis sizes for the the aortic annulus based on MDCT images. &e method has Model-observer 1 Model-observer 1 Observer 2-observer 1 Observer 2-observer 1 6 Journal of Interventional Cardiology Edwards sapien 3 Medtronic evolut TAVR 3 4 5 2 Model Observer 1 Observer 2 Model Observer 1 Observer 2 20 26 23 29 (a) (b) Figure 6: &e agreement between prosthesis sizes from the Edwards Sapien 3 (a) and Medtronic Evolut TAVR sizing chart (b). &e plots represent how many sizes were selected for each available device size based on the model, observer 1, and observer 2. &e arrows between the plots indicate disagreement with observer 1 (under- or overestimation). &e weights indicate the number of patients that were sized differently as compared to observer 1. Edwards Sapien 3 and the Medtronic Evolute device. &e &e method can detect the area and perimeter from the automatically determined measurements result in device aortic annular plane within seconds, which may have an impact on reducing operator analysis time and errors in size selections that accord well with the device sizes selected by operator 1 based on his measurements, which again an exponentially growing market. If this method was confirms the adequate model accuracy. &e total analysis combined with an automatic aortic annular plane de- time from aortic annular plane to prothesis size is below 1 tection method, the overall time reduction would be second. considerable. In addition to a time reduction of analysis In the literature, similar studies have been conducted. and, thus, procedure planning, the physician saves time as Queiros ´ et al. proposed a method for detecting the correct he/she is liberated from this planning/analysis. Also, the TAVI prosthesis size from the aortic valve annulus area analysis concerns an independent automated process that using aortic segmentation and statistical shape models [12]. will enhance the output quality. Reduced overall TAVI &eir full-automatic approach detected 92% of the prosthesis costs may be obtained by embedding the method in sizes and their semiautomatic approach 100%. &is single- software that allows manual corrections (e.g., to correct outliers). &is embedding could also yield a continuous center study included 104 patients with a severe degree of calcification, mitral valve prosthesis, and pacemakers. &e learning platform where the data of a new patient, vali- authors introduced an overlapping area of 35 mm dated by an expert, can be added to the training dataset, and 40 mm between the 3 available prosthesis sizes of the thus improving future detections. Edwards Sapien 3 and XT. Unfortunately, this overlapping Although the presented method has proven to be reli- area makes it difficult to assess the true predictive power of able, there are a few limitations related to the current ap- the method and to compare with our results. Also, the final proach. In a few cases, relatively large differences remain processing time was not reported in this study. between the predicted area from our model and that from an Our presented method is based on a different tech- individual human observer. Compared to observer 1, the nique and goes, in our opinion, a step further than the largest overestimation of our model amounts to 10% and the work described in [12]. Our study includes both aortic largest underestimation to 9%. However, in those cases, observer 2 tended to agree with the predicted value (1% valve annulus perimeter and area; therefore, the pros- thesis size selection can be expanded to perimeter as well difference between observer 2 and the model). &is may indicate that the model has generalized beyond the ground as area dependent devices. Next, multicenter data were used for training and validating the model, which may truth; in other words, it has learned to look beyond the few indicate robustness to unknown centers. No overlapping inaccuracies of its teacher. &e maximum difference between region was used in order to follow the manufacturers’ the predicted perimeter and observer 1 was the same patient guidelines and leave the final interpretation of the output as the areas maximum difference (with a 7% over- of the method to the physician. Finally, the processing estimation). &e minimum difference between predicated time is around one second per patient, which makes the area and observer 1 was a 5% underestimation (a 3 mm difference). method fast. Selected sizes Selected sizes Journal of Interventional Cardiology 7 It should be noted that the proposed method is not a were similarly obtained as the training details. (Supple- TAVI planning tool, nor does it intend to replace the mentary Materials) interventional cardiologist. &ere are other measurements required for the planning of a TAVI which are not included References in this study. &ese measurements include the distance from [1] C. R. Smith, M. B. Leon, M. J. Mack et al., “Transcatheter the aortic annular plane to the ostium of the coronary ar- versus surgical aortic-valve replacement in high-risk patients,” teries, the area of Sinus of Valsalva, sinotubular junction, New England Journal of Medicine, vol. 364, no. 23, and others and will be addressed in future work. It would pp. 2187–2198, 2011. also be interesting to measure the impact of this method [2] M. B. Leon, C. R. Smith, M. J. Mack et al., “Transcatheter or prospectively. surgical aortic-valve replacement in intermediate-risk pa- tients,” New England Journal of Medicine, vol. 374, no. 17, pp. 1609–1620, 2016. 4.Conclusions [3] M. J. Reardon, N. M. Van Mieghem, J. J. Popma et al., “Surgical or transcatheter aortic-valve replacement in in- In conclusion, this study shows that automated TAVI device termediate-risk patients,” New England Journal of Medicine, size selection using the proposed method is fast, accurate, vol. 376, no. 14, pp. 1321–1331, 2017. and reproducible. Comparison with the interobserver var- [4] J. J. Popma, G. M. Deeb, S. J. Yakubov et al., “Transcatheter iability has shown the reliability of the strategy, and em- aortic-valve replacement with a self-expanding valve in low- bedding this tool based on deep learning in the preoperative risk patients,” New England Journal of Medicine, vol. 380, planning routine has the potential to increase the efficiency no. 18, pp. 1706–1715, 2019. while ensuring accuracy. [5] M. J. Mack, M. B. Leon, V. H. &ourani et al., “Transcatheter aortic-valve replacement with a balloon-expandable valve in low-risk patients,” New England Journal of Medicine, vol. 380, Data Availability no. 18, pp. 1695–1705, 2019. [6] A. P. Durko, R. L. Osnabrugge, N. M. Van Mieghem et al., &e statistical data used to support the findings of this study “Annual number of candidates for transcatheter aortic valve are available from the corresponding author upon request. implantation per country: current estimates and future pro- &e anonymized image data used to support the findings of jections,” European Heart Journal, vol. 39, no. 28, pp. 2635– this study were supplied by FEops N.V. under license and so 2642, 2018. cannot be made freely available. [7] A. M. Kasel, S. Cassese, S. Bleiziffer et al., “Standardized imaging for aortic annular sizing,” JACC: Cardiovascular Conflicts of Interest Imaging, vol. 6, no. 2, pp. 249–262, 2013. [8] O. Ronneberger, P. Fischer, and T. Brox, “U-net: convolu- Peter de Jaegere is a consultant for Medtronic. Johan Bos- tional networks for biomedical image segmentation,” Lecture mans is a consultant for Medtronic. Ole De Backer has been Notes in Computer Science, vol. 9351, pp. 234–241, 2015. [9] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning a consultant for Abbott. Matthieu De Beule and Peter for image recognition,” in Proceedings of the IEEE Computer Mortier are shareholders of FEops. Joni Dambre and Pat- Society Conference on Computer Vision and Pattern Recog- ricio Astudillo have no conflicts of interest to declare. nition, pp. 770–778, Patna, India, December 2016. [10] J. Canny, “A computational approach to edge detection,” IEEE Acknowledgments Transactions on Pattern Analysis and Machine Intelligence, vol. PAMI-8, no. 6, pp. 679–698, 1986. We want to thank the two operators for all their work. All [11] A. Paszke, S. Gross, S. Chintala et al., Automatic Differenti- computational work was performed on Nvidia GPUs from ation in PyTorch, in Proceedings of the 31st Conference on the Nvidia GPU Grant Program. &is work was supported by Neural Information Processing System, Long Beach, CA, USA, the European Commission within the Horizon 2020 December 2017. Framework through the MSCA-ITN-ETN European [12] S. Queiros, ´ C. Dubois, P. Morais et al., “Automatic 3D aortic annulus sizing by computed tomography in the planning of Training Networks (project no. 642458). transcatheter aortic valve implantation,” Journal of Cardio- vascular Computed Tomography, vol. 11, no. 1, pp. 25–32, Supplementary Materials Figure 1: &e overview of the deep learning architecture. Figure 2: &e overview of the residual block: the input is expanded to the desired number of filters with a convolu- tional 2D layer with kernel size 1. After a sequence of convolutional layers with kernel size 3, batch normalization [3], and ReLU activation function [4], the output is summed with the output of the first convolutional layer followed by a final ReLU activation function. Table 1: Training details. All hyperparameters were obtained by performing k-fold cross- validation on the training set (with k = 5) and a fixed random seed. Table 2: Data-augmentation details. All parameters MEDIATORS of INFLAMMATION The Scientific Gastroenterology Journal of World Journal Research and Practice Diabetes Research Disease Markers Hindawi Hindawi Publishing Corporation Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 http://www www.hindawi.com .hindawi.com V Volume 2018 olume 2013 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 International Journal of Journal of Immunology Research Endocrinology Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Submit your manuscripts at www.hindawi.com BioMed PPAR Research Research International Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Journal of Obesity Evidence-Based Journal of Journal of Stem Cells Complementary and Ophthalmology International Alternative Medicine Oncology Hindawi Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2013 Parkinson’s Disease Computational and Behavioural Mathematical Methods AIDS Oxidative Medicine and in Medicine Neurology Research and Treatment Cellular Longevity Hindawi Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Interventional Cardiology Hindawi Publishing Corporation

Enabling Automated Device Size Selection for Transcatheter Aortic Valve Implantation

Download PDF
8 pages

Loading next page...
 
/lp/hindawi-publishing-corporation/enabling-automated-device-size-selection-for-transcatheter-aortic-Ba8iR0CZBd

References

References for this paper are not available at this time. We will be adding them shortly, thank you for your patience.

Publisher
Hindawi Publishing Corporation
Copyright
Copyright © 2019 Patricio Astudillo et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
ISSN
1540-8183
eISSN
0896-4327
DOI
10.1155/2019/3591314
Publisher site
See Article on Publisher Site

Abstract

Hindawi Journal of Interventional Cardiology Volume 2019, Article ID 3591314, 7 pages https://doi.org/10.1155/2019/3591314 Research Article Enabling Automated Device Size Selection for Transcatheter Aortic Valve Implantation 1,2 1 3 4 5 Patricio Astudillo , Peter Mortier, Johan Bosmans, Ole De Backer, Peter de Jaegere, 1 2 Matthieu De Beule, and Joni Dambre FEops, Technologiepark-Zwijnaarde 122, Ghent, Belgium Department of Electronics and Information Systems, UGent-imec, Technologiepark-Zwijnaarde 126, Ghent, Belgium University Hospital Antwerp (UZA), Antwerp, Belgium Department of Cardiology, Rigshospitalet University Hospital, Copenhagen, Denmark Department of Cardiology, Erasmus MC, Rotterdam, Netherlands Correspondence should be addressed to Patricio Astudillo; patricio.astudillo@feops.com Received 29 April 2019; Revised 5 August 2019; Accepted 16 August 2019; Published 3 November 2019 Academic Editor: Amit Segev Copyright © 2019 Patricio Astudillo et al. �is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. �e number of transcatheter aortic valve implantation (TAVI) procedures is expected to increase signi‡cantly in the coming years. Improving eˆciency will become essential for experienced operators performing large TAVI volumes, while new operators will require training and may bene‡t from accurate support. In this work, we present a fast deep learning method that can predict aortic annulus perimeter and area automatically from aortic annular plane images. We propose a method combining two deep convolutional neural networks followed by a postprocessing step. �e models were trained with 355 patients using modern deep learning techniques, and the method was evaluated on another 118 patients. �e method was validated against an interoperator variability study of the same 118 patients. �e di‘erences between the manually obtained aortic annulus measurements and the automatic predictions were similar to the di‘erences between two independent observers (paired di‘. of 3.3 ± 16.8 mm vs. 1.3 ± 21.1 mm for the area and a paired di‘. of 0.6 ± 1.7 mm vs. 0.2 ± 2.5 mm for the perimeter). �e area and perimeter were used to retrieve the suggested prosthesis sizes for the Edwards Sapien 3 and the Medtronic Evolut device retrospectively. �e au- tomatically obtained device size selections accorded well with the device sizes selected by operator 1. �e total analysis time from aortic annular plane to prosthesis size was below one second. �is study showed that automated TAVI device size selection using the proposed method is fast, accurate, and reproducible. Comparison with the interobserver variability has shown the reliability of the strategy, and embedding this tool based on deep learning in the preoperative planning routine has the potential to increase the eˆciency while ensuring accuracy. �e number of TAVI procedures is increasing each year 1. Introduction rapidly [6], and considering the recent clinical data for low- Transcatheter aortic valve implantation (TAVI) has become risk patients will lead to an accelerated expansion in the the preferred treatment for patients with aortic stenosis at coming years. As a result, scalability of the complete pro- high risk for surgical aortic valve replacement (SAVR) [1]. cedure, including preoperative planning, becomes an im- Recently, it was concluded that, for intermediate-risk pa- portant aspect. Experienced operators can enlarge their tients, TAVI was similar to SAVR with respect to the pri- volume of TAVI cases, for example, by increasing pro- mary end-point of death or disabling stroke [2, 3]. Very cedural eˆciency. On the other hand, many new operators recent clinical data even show that TAVI is at least as good as will need to be trained, which logically leads to increased SAVR in low-risk patients [4, 5]. risks due to their limited experience. When focusing on the 2 Journal of Interventional Cardiology teach the neural network how to segment the aortic annular preoperative planning, accurate automated detection of the aortic annulus dimensions directly from multidetector plane using the ground truth annotations of the aortic annulus (Figure 1). computed tomography (MDCT) images could not only increase efficiency but also at the same time reduce operator &e deep learning model requires an architecture in variability, thereby minimizing the impact of experience on order to process the resampled and clipped aortic annular TAVI sizing. planes and compare the output of the model with the In this work, we present a deep learning method that can binary masks. &e used architecture was inspired by U-Net predict the aortic annulus perimeter and area automatically. [8] and deep residual nets [9] and consisted of two paths: a &e method is validated against an interoperator variability downscaling and an upscaling path. &e downscaling path study to assess its accuracy. As a final step, the impact of the extracted information from the aortic annular plane, and proposed method on the prosthesis size selection for both the upscaling path translated this information into a seg- mented aortic annulus. &e final sigmoid activation the Edwards Lifesciences and Medtronic transcatheter aortic bioprostheses was evaluated. function ensured that the output of the model contained probability values. &e details of the deep learning archi- tecture, training, and data-augmentation techniques are 2.Materials and Methods given in appendix A in Supplementary Materials (available 2.1. MDCT Imaging. &is retrospective study used the here). anonymized data of 473 patients collected from multiple &e deep learning architecture was used during the centra. &e mean age of this cohort was 80.82± 7.18 years, training phase to teach a deep learning model to segment the and 55% of the patients were female. &ere were 36 bicuspid aortic annulus from the aortic annular plane. patients in this cohort. &e patient data consisted of volu- metric MDCT images which were acquired to plan a TAVI procedure. &erefore, all MDCT images were contrast-en- 2.3.1. Training. Two models were trained using the training hanced and contained a certain degree of aortic stenosis. &e dataset and validated with the validation dataset. One model average row, column, and slice thickness of the MDCT was trained for each of the two resolutions (1 mm and images were 512.05 mm, 511.85 mm, and 0.83 mm. &e 0.5 mm) of the aortic annular planes. &e validation dataset aortic annular planes (AAP) were manually identified from consisted of the same 118 patients that were used for the the volumetric MDCT images using the standard method [7] interobserver variability study, and the training dataset and were used as input for this study. For this retrospective consisted of the remaining 355 patients. &e 36 bicuspid study, formal consent is not required. patients were distributed equally over the training and validation datasets. 2.2.ManualDetection. &e border of the aortic annulus was manually identified from the aortic annular planes by ob- 2.4.Detection. After training one model for each resolution, server 1. &e data of observer 1 were considered the ground a detection strategy was used to combine the output of both truth in this study. Observer 2 repeated this for 118 ran- models and to derive patient-specific anatomical in- domly selected patients in order to assess the interoperator formation: the area and perimeter of the aortic annulus. variability. Both observers applied the same manual method, &e detection of the area and perimeter of the aortic which consists of visual detection of the aortic annulus annulus of a single patient was performed in two steps: a within the AAP and annotating it using Mimics Innovation deep learning step and a postprocessing step. During the Suite 18 (Materialise, Leuven, Belgium). deep learning step, the aortic annular planes were ana- lysed by both models, and the output was combined and 2.3.AutomaticDetection. &is study aims at automating the normalized to a probability output that identified the manual segmentation and derives clinical patient-specific region of interest. During the postprocessing step, the contour of the region of interest was located with canny information as a postprocessing step. Preprocessing of the ground truth images and aortic annulus annotations were edge detection [10] from the probability output. &e area and perimeter were derived from this contour and serve necessary in order to prepare the data for training the deep learning models. as the final predicted output of the detection phase &e aortic annular planes were clipped and resampled in (Figure 2). order to fit the neural networks’ input. &e aortic annular As a final step, the derived aortic annulus dimensions were planes were resampled to an isotropic 1 mm resolution. As used to assess correct prosthesis size. &e perimeter was used to the deep learning network expected a 128 ×128 pixel plane select the proper Medtronic Evolut TAVR prosthesis size as input, the resampled aortic annular planes were clipped (https://www.medtronic.com/content/dam/medtronic-com/ around the center of the aortic annulus. A second isotropic products/cardiovascular/heart-valves-transcatheter/corevalve- 0.5 mm resolution was generated and clipped in the same evolut-r/documents/201709637EN-Evolut-PRO-TAV-in- SAV-Interactive-Sizing-Guide-FINAL.pdf), and the area was manner in order to double the level of segmentation detail. Cubic spline interpolation was used in order to retain the used to identify the Edwards Sapien 3 prosthesis size (https:// www.accessdata.fda.gov/cdrh_docs/pdf14/P140031c.pdf) sim- original Hounsfield units in the resampled aortic annular planes (Figure 1). Binary masks were generated in order to ilar to the manufacturer’s sizing matrix used in clinical practice. Journal of Interventional Cardiology 3 (a) (b) (c) (d) Figure 1: Examples of the aortic annular plane and the accompanying binary masks. &e resampled and clipped aortic annular planes (a) and the binary masks (b) with different resolutions, 1.0 mm (c) and 0.5 mm (d). 2.5.StatisticalAnalysis. &e Shapiro–Wilk test was performed patients for validation and observer variability assessment, it to test for normal distribution, and none of the predicted was possible to compare the method with both observers. distributions were normally distributed. Pearson correlation &e detection phase consisted of a deep learning phase coefficient was computed to evaluate the correlation between and a postprocessing phase. &e deep learning phase was model and both observers (with excellent correlation R > 0.9). validated by comparing the predicted segmentation (model) &e agreement between manual and the automatic landmark with the segmentation of both observers using the dice co- locations were evaluated using the nonparametric signed efficient. &e mean Dice score between model and observer 1 Wilcoxon test (with a significant p value <0.001). Bland– was 96% whereas the mean Dice score between both model Altman analysis for area and perimeter between model and and observer 2 and observer 1 and 2 was 89%. &e higher observer 1 and between both observers was performed. mean Dice score between model and observer 1 is expected because the model was trained with the data from observer 1. &e postprocessing phase derived the area and perimeter 2.6. Implementation. All the computational work was per- from the predicted segmentation and was validated by formed on a multicore computer with Titan X and P6000 comparing the predicted area and perimeter with the area GPUs (NVIDIA Corporation, Los Alamitos, CA). &e and perimeter of both observers. When comparing the models and the deep learning pipeline were developed with predicted anatomical measurements of the model with the PyTorch v0.4.1 [11]. data of both observers, there was no significant difference between the model and both observers for the area mea- 3.Results and Discussion surements. &e mean paired difference for all measurements 3.1. Results. &e proposed method trained two models, and was around zero, which means that the predicted anatomical the detection phase was validated using the 118 patients used measurements could be used in the same manner as the in the interoperator variability study. By using the same output of observer 1 or 2 (Table 1). 4 Journal of Interventional Cardiology Predict Compare Figure 2: A general overview of the method: the model predicts the probability plane from the original aortic annular plane. &e contours are detected, and the predicted area and perimeter are compared with the ground truth (GT). 2 Perimeter (mm) Area (mm ) 60 70 80 90 100 110 300 400 500 600 700 Observer 1 Observer 1 2 Model (R = 0.97) Model (R = 0.98) 2 Observer 2 (R = 0.94) Observer 2 (R = 0.97) (a) (b) Figure 3: Scatter plots comparing the interobserver correlation for the area (a) and perimeter (b). Excellent correction values were obtained between method’s ability to predict the correct prosthesis size model and observer 1 for the area (0.98) and perimeter (compared to both observers). &e predicted area and pe- (0.97). &e correlation values between observer 1 and 2 for rimeter were used to retrieve the Edwards Sapien 3 and the area (0.97) and perimeter (0.94) indicate that the manual Medtronic Evolut TAVR prosthesis sizes. &e automatically method is accurate (Figure 3). selected valve sizes were compared with valve sizes resulting Bland–Altman plots of the predicted and measured (ob- from the annular measurements of both observers. &e ratio server 1) area and perimeter are depicted in Figures 4 and 5. of agreement for Edwards Sapien 3 between model and both It is worth noting, when interpreting the Bland–Altman plots, observers is almost equal: 0.86 between model and observer that the model was repeatable since consecutive predictions 1 and 0.88 between both observers. &e ratio of agreement per patient yielded the same output. for the Medtronic Evolut TAVR prosthesis sizes between &e validation of the segmentation abilities and the area model and both observers is similar: 0.89 between model and and perimeter assessment were required to validate the observer 1 and 0.86 between both observers (Figure 6). Postprocess Model/observer 2 Model/observer 2 Journal of Interventional Cardiology 5 Table 1: A comparison of the anatomical measurements between model and both observers. Model vs. observer 1 Model vs. observer 2 Observer 1 vs. observer 2 Paired diff. p value Paired diff. p value Paired diff. p value 2) Area (mm 3.3± 16.8 0.008 2.0± 22.4 0.046 1.3± 21.1 0.752 Perimeter (cm) 0.6± 1.7 0.0001 0.5± 2.6 0.0016 0.2± 2.5 0.513 Paired difference reported as mean± standard deviation. 2 2 Area (mm ) Area (mm ) 100 100 50 50 0 0 –50 –50 –100 –100 300 400 500 600 700 300 400 500 600 700 Average of model and observer 1 Average of observer 2 and observer 1 (a) (b) Figure 4: Bland–Altman plots comparing the aortic annulus area for model vs. observer 1 (a) and both observers (b). Perimeter (mm) Perimeter (mm) 10 10 5 5 0 0 –5 –5 –10 –10 60 65 70 75 80 85 90 95 100 60 65 70 75 80 85 90 95 100 Average of model and observer 1 Average of observer 2 and observer 1 (a) (b) Figure 5: Bland–Altman plots comparing the aortic annulus perimeter for model vs. observer 1 (a) and both observers (b). Finally, it is relevant to report the processing time of the been validated on 118 patients to evaluate its accuracy, and manual and automated methods. &e automatic processing the results show that the area and perimeter can be predicted time from aortic annular plane to segmentation, anatomical in an automatic, reproducible, fast, and accurate way by measurement, and prosthesis size is below 1 second. combining the results of two networks followed by a postprocessing step. &e differences between the manually obtained aortic annulus measurements and the automatic 3.2. Discussion. In this work, an automated method is predictions are similar to the differences between two in- dependent observers, which indicates a satisfying accuracy proposed to facilitate and optimize the preoperative TAVI planning. It automatically predicts the area and perimeter of of the proposed approach. &e area and perimeter have also been used to retrieve the suggested prosthesis sizes for the the aortic annulus based on MDCT images. &e method has Model-observer 1 Model-observer 1 Observer 2-observer 1 Observer 2-observer 1 6 Journal of Interventional Cardiology Edwards sapien 3 Medtronic evolut TAVR 3 4 5 2 Model Observer 1 Observer 2 Model Observer 1 Observer 2 20 26 23 29 (a) (b) Figure 6: &e agreement between prosthesis sizes from the Edwards Sapien 3 (a) and Medtronic Evolut TAVR sizing chart (b). &e plots represent how many sizes were selected for each available device size based on the model, observer 1, and observer 2. &e arrows between the plots indicate disagreement with observer 1 (under- or overestimation). &e weights indicate the number of patients that were sized differently as compared to observer 1. Edwards Sapien 3 and the Medtronic Evolute device. &e &e method can detect the area and perimeter from the automatically determined measurements result in device aortic annular plane within seconds, which may have an impact on reducing operator analysis time and errors in size selections that accord well with the device sizes selected by operator 1 based on his measurements, which again an exponentially growing market. If this method was confirms the adequate model accuracy. &e total analysis combined with an automatic aortic annular plane de- time from aortic annular plane to prothesis size is below 1 tection method, the overall time reduction would be second. considerable. In addition to a time reduction of analysis In the literature, similar studies have been conducted. and, thus, procedure planning, the physician saves time as Queiros ´ et al. proposed a method for detecting the correct he/she is liberated from this planning/analysis. Also, the TAVI prosthesis size from the aortic valve annulus area analysis concerns an independent automated process that using aortic segmentation and statistical shape models [12]. will enhance the output quality. Reduced overall TAVI &eir full-automatic approach detected 92% of the prosthesis costs may be obtained by embedding the method in sizes and their semiautomatic approach 100%. &is single- software that allows manual corrections (e.g., to correct outliers). &is embedding could also yield a continuous center study included 104 patients with a severe degree of calcification, mitral valve prosthesis, and pacemakers. &e learning platform where the data of a new patient, vali- authors introduced an overlapping area of 35 mm dated by an expert, can be added to the training dataset, and 40 mm between the 3 available prosthesis sizes of the thus improving future detections. Edwards Sapien 3 and XT. Unfortunately, this overlapping Although the presented method has proven to be reli- area makes it difficult to assess the true predictive power of able, there are a few limitations related to the current ap- the method and to compare with our results. Also, the final proach. In a few cases, relatively large differences remain processing time was not reported in this study. between the predicted area from our model and that from an Our presented method is based on a different tech- individual human observer. Compared to observer 1, the nique and goes, in our opinion, a step further than the largest overestimation of our model amounts to 10% and the work described in [12]. Our study includes both aortic largest underestimation to 9%. However, in those cases, observer 2 tended to agree with the predicted value (1% valve annulus perimeter and area; therefore, the pros- thesis size selection can be expanded to perimeter as well difference between observer 2 and the model). &is may indicate that the model has generalized beyond the ground as area dependent devices. Next, multicenter data were used for training and validating the model, which may truth; in other words, it has learned to look beyond the few indicate robustness to unknown centers. No overlapping inaccuracies of its teacher. &e maximum difference between region was used in order to follow the manufacturers’ the predicted perimeter and observer 1 was the same patient guidelines and leave the final interpretation of the output as the areas maximum difference (with a 7% over- of the method to the physician. Finally, the processing estimation). &e minimum difference between predicated time is around one second per patient, which makes the area and observer 1 was a 5% underestimation (a 3 mm difference). method fast. Selected sizes Selected sizes Journal of Interventional Cardiology 7 It should be noted that the proposed method is not a were similarly obtained as the training details. (Supple- TAVI planning tool, nor does it intend to replace the mentary Materials) interventional cardiologist. &ere are other measurements required for the planning of a TAVI which are not included References in this study. &ese measurements include the distance from [1] C. R. Smith, M. B. Leon, M. J. Mack et al., “Transcatheter the aortic annular plane to the ostium of the coronary ar- versus surgical aortic-valve replacement in high-risk patients,” teries, the area of Sinus of Valsalva, sinotubular junction, New England Journal of Medicine, vol. 364, no. 23, and others and will be addressed in future work. It would pp. 2187–2198, 2011. also be interesting to measure the impact of this method [2] M. B. Leon, C. R. Smith, M. J. Mack et al., “Transcatheter or prospectively. surgical aortic-valve replacement in intermediate-risk pa- tients,” New England Journal of Medicine, vol. 374, no. 17, pp. 1609–1620, 2016. 4.Conclusions [3] M. J. Reardon, N. M. Van Mieghem, J. J. Popma et al., “Surgical or transcatheter aortic-valve replacement in in- In conclusion, this study shows that automated TAVI device termediate-risk patients,” New England Journal of Medicine, size selection using the proposed method is fast, accurate, vol. 376, no. 14, pp. 1321–1331, 2017. and reproducible. Comparison with the interobserver var- [4] J. J. Popma, G. M. Deeb, S. J. Yakubov et al., “Transcatheter iability has shown the reliability of the strategy, and em- aortic-valve replacement with a self-expanding valve in low- bedding this tool based on deep learning in the preoperative risk patients,” New England Journal of Medicine, vol. 380, planning routine has the potential to increase the efficiency no. 18, pp. 1706–1715, 2019. while ensuring accuracy. [5] M. J. Mack, M. B. Leon, V. H. &ourani et al., “Transcatheter aortic-valve replacement with a balloon-expandable valve in low-risk patients,” New England Journal of Medicine, vol. 380, Data Availability no. 18, pp. 1695–1705, 2019. [6] A. P. Durko, R. L. Osnabrugge, N. M. Van Mieghem et al., &e statistical data used to support the findings of this study “Annual number of candidates for transcatheter aortic valve are available from the corresponding author upon request. implantation per country: current estimates and future pro- &e anonymized image data used to support the findings of jections,” European Heart Journal, vol. 39, no. 28, pp. 2635– this study were supplied by FEops N.V. under license and so 2642, 2018. cannot be made freely available. [7] A. M. Kasel, S. Cassese, S. Bleiziffer et al., “Standardized imaging for aortic annular sizing,” JACC: Cardiovascular Conflicts of Interest Imaging, vol. 6, no. 2, pp. 249–262, 2013. [8] O. Ronneberger, P. Fischer, and T. Brox, “U-net: convolu- Peter de Jaegere is a consultant for Medtronic. Johan Bos- tional networks for biomedical image segmentation,” Lecture mans is a consultant for Medtronic. Ole De Backer has been Notes in Computer Science, vol. 9351, pp. 234–241, 2015. [9] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning a consultant for Abbott. Matthieu De Beule and Peter for image recognition,” in Proceedings of the IEEE Computer Mortier are shareholders of FEops. Joni Dambre and Pat- Society Conference on Computer Vision and Pattern Recog- ricio Astudillo have no conflicts of interest to declare. nition, pp. 770–778, Patna, India, December 2016. [10] J. Canny, “A computational approach to edge detection,” IEEE Acknowledgments Transactions on Pattern Analysis and Machine Intelligence, vol. PAMI-8, no. 6, pp. 679–698, 1986. We want to thank the two operators for all their work. All [11] A. Paszke, S. Gross, S. Chintala et al., Automatic Differenti- computational work was performed on Nvidia GPUs from ation in PyTorch, in Proceedings of the 31st Conference on the Nvidia GPU Grant Program. &is work was supported by Neural Information Processing System, Long Beach, CA, USA, the European Commission within the Horizon 2020 December 2017. Framework through the MSCA-ITN-ETN European [12] S. Queiros, ´ C. Dubois, P. Morais et al., “Automatic 3D aortic annulus sizing by computed tomography in the planning of Training Networks (project no. 642458). transcatheter aortic valve implantation,” Journal of Cardio- vascular Computed Tomography, vol. 11, no. 1, pp. 25–32, Supplementary Materials Figure 1: &e overview of the deep learning architecture. Figure 2: &e overview of the residual block: the input is expanded to the desired number of filters with a convolu- tional 2D layer with kernel size 1. After a sequence of convolutional layers with kernel size 3, batch normalization [3], and ReLU activation function [4], the output is summed with the output of the first convolutional layer followed by a final ReLU activation function. Table 1: Training details. All hyperparameters were obtained by performing k-fold cross- validation on the training set (with k = 5) and a fixed random seed. Table 2: Data-augmentation details. All parameters MEDIATORS of INFLAMMATION The Scientific Gastroenterology Journal of World Journal Research and Practice Diabetes Research Disease Markers Hindawi Hindawi Publishing Corporation Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 http://www www.hindawi.com .hindawi.com V Volume 2018 olume 2013 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 International Journal of Journal of Immunology Research Endocrinology Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Submit your manuscripts at www.hindawi.com BioMed PPAR Research Research International Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Journal of Obesity Evidence-Based Journal of Journal of Stem Cells Complementary and Ophthalmology International Alternative Medicine Oncology Hindawi Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2013 Parkinson’s Disease Computational and Behavioural Mathematical Methods AIDS Oxidative Medicine and in Medicine Neurology Research and Treatment Cellular Longevity Hindawi Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Journal

Journal of Interventional Cardiology – Hindawi Publishing Corporation

Published: Nov 3, 2019

References

  • Transcatheter versus surgical aortic-valve replacement in high-risk patients,
    C. R. Smith, M. B. Leon, M. J. Mack et al.
  • Transcatheter or surgical aortic-valve replacement in intermediate-risk patients,
    M. B. Leon, C. R. Smith, M. J. Mack et al.
  • Surgical or transcatheter aortic-valve replacement in intermediate-risk patients,
    M. J. Reardon, N. M. Van Mieghem, J. J. Popma et al.
  • Transcatheter aortic-valve replacement with a self-expanding valve in low-risk patients,
    J. J. Popma, G. M. Deeb, S. J. Yakubov et al.
  • Transcatheter aortic-valve replacement with a balloon-expandable valve in low-risk patients,
    M. J. Mack, M. B. Leon, V. H. Thourani et al.
  • Annual number of candidates for transcatheter aortic valve implantation per country: current estimates and future projections,
    A. P. Durko, R. L. Osnabrugge, N. M. Van Mieghem et al.
  • Standardized imaging for aortic annular sizing,
    A. M. Kasel, S. Cassese, S. Bleiziffer et al.
  • U-net: convolutional networks for biomedical image segmentation,
    O. Ronneberger, P. Fischer, and T. Brox
  • Deep residual learning for image recognition,
    K. He, X. Zhang, S. Ren, and J. Sun
  • A computational approach to edge detection,
    J. Canny
  • Automatic 3D aortic annulus sizing by computed tomography in the planning of transcatheter aortic valve implantation,
    S. Queirós, C. Dubois, P. Morais et al.