Roughly Collected Dataset for Contact Force Sensing Catheter
Seunghyuk Cho, Minsoo Koo, Dongwoo Kim, Juyong Lee, Yeonwoo Jung, Kibyung Nam, Changmo Hwang
RRoughly Collected Dataset for Contact Force Sensing Catheter
Seunghyuk Cho ∗† , Minsoo Koo † , Dongwoo Kim, Juyoung Lee, Yeonwoo Jung, Kibyung Nam, Changmo Hwang ‡ Abstract — With rise of interventional cardiology, CatheterAblation Therapy (CAT) has established itself as a first-line solution to treat cardiac arrhythmia. Although CAT isa promising technique, cardiologist lacks vision inside thebody during the procedure, which may cause serious clinicalsyndromes. To support accurate clinical procedure, ContactForce Sensing (CFS) system is developed to find a position ofthe catheter tip through the measure of contact force betweencatheter and heart tissue. However, the practical usability ofcommercialized CFS systems is not fully understood due toinaccuracy in the measurement. To support the development ofmore accurate system, we develop a full pipeline of CFS systemwith newly collected benchmark dataset through a contact forcesensing catheter in simplest hardware form. Our dataset wasroughly collected with human noise to increase data diversity.Through the analysis of the dataset, we identify a problemdefined as Shift of Reference (SoR), which prevents accuratemeasurement of contact force. To overcome the problem, weconduct the contact force estimation via standard deep neuralnetworks including for Recurrent Neural Network (RNN), FullyConvolutional Network (FCN) and Transformer. An averageerror in measurement for RNN, FCN and Transformer are,respectively, 2.46g, 3.03g and 3.01g. Through these studies, wetry to lay a groundwork, serve a performance criteria for futureCFS system research and open a publicly available dataset topublic.
I. INTRODUCTIONCatheter Ablation Therapy (CAT) is a clinical method totreat arrhythmia, a series of irregular heartbeat caused bymalfunctioning of certain tissues in heart. Depending on theheart rhythm, arrhythmia can be classified into tachycardia,bradycardia and fibrillation. In the past, anti-arrhythmicdrugs was the major treatment for arrhythmia. However,CAT now becomes the first-line treatment for arrhythmia.Comparing to the anti-arrhythmic drugs, CAT significantlyreduces the chance of pulmonary vein re-connection. Forexample, Amiodarone can only temporarily relieve illness
This work was supported by the Technology Innovation Program (or In-dustrial Strategic Technology Development Program) (Grant No.: 20008028)funded By the Ministry of Trade, Industry & Energy (MOTIE).* This work was done while S.H. Cho was research intern in AsanMedical Center † Equal Contribution ‡ Corresponding AuthorS. H. Cho, M. S. Koo, K. B. Nam and C. M. Hwang arewith Asan Medical Center, Seoul, 05505, Korea. (corresponding au-thor to provide phone: 82-2-3010-4097; fax: 82-2-3010-4182; email: [email protected] )S. H. Cho, D. W. Kim, J. Y. Lee and Y. W. Jeong are with PohangUniversity of Science and Technology, Pohang, 37673, Korea. (email: [email protected] )M. S. Koo is with Xi’an Jiaotong University Health Science Center, Xi’an,710061, China. (email: [email protected] )K. B. Nam and C. M Hwang are with University of Ulsan College ofMedicine, 05505, Korea. (email: [email protected] ) Normal Impulse Conduction Pulmonary VeinLeft Atrium Treated AreasMapping CatheterAblation CatheterFig. 1. Illustration shows how ablation catheter is guided to dysrhythmicfocus. It starts with the cardiologist inserting a 5-8 Fr wide threadlikepipe called catheter into an artery or vein through groin or arm. Then thecardiologist slowly navigate the catheter until it reaches to the left atrium.If cardiologist successfully pinpoint the dysrhythmic focus and locate thecatheter on it, she or he induces Radio-Frequency (RF) energy to impairthe problematic cardiac tissue. The procedure is complicated and it is hardfor cardiologist to know if the procedure is going well or not. of patient [1]. Patients need to take the drug often for a longperiod to see effects similar to CAT. Moreover, comparedto classic surgical procedures, patients do not need to worryabout inherent aesthetic, clinical problems [2]. Interventionalmethods also offers lower risk of infection, less blood loss,and fast recovery time for patients.CAT has many benefits for arrhythmia treatment but itsmechanism is complex. Fig. 1 illustrates the procedure ofCAT for Radio-Frequency Ablation (RFA) treatment. InCAT, even a trivial miss in procedure may lead to seriouscomplications. Possibility of cardiac perforation, effusionand tamponade is still 0.98% per procedure and 1.46% perpatient [3]. X-Ray radiography and electrodes on catheterare developed to support the procedure, however, they arenot enough for ideal procedure. Contact Force Sensing(CFS) system is developed to solve these problems. It usesthreadlike force sensor to measure the contact force ofcardiac tissue - catheter contact made with Fiber Bragg a r X i v : . [ c s . R O ] F e b rating (FBG). FBG is a phase grating inscribed to the cable.Gating period and length, core refractive index determinewhether the grating has a high or low reflectivity over a wideor narrow range of wavelengths [4]. There are number ofmanuscripts about implementation of contact force sensingcatheter with FBGs [5] [6]. Fig. 2 shows the idea of howFBGs works as a force sensor.According to [7], it is known that CAT reduces the timeof inducing RF and fluoroscopy significantly. Through thesestudies, we conclude that CFS is an indispensable systemin CAT. The fact the system is so sensitive to accuracymean it should be studied continuously. However, most ofthem only reports their performance without implementationand measurement details. For example, TactiCath Quartz(TCQ) ablation catheter (Abbott Laboratories, Abbott Park,IL, USA) [8][9] and ThermoCool SmartTouch® Catheter(Johnson & Johnson, New Brunswick, NJ, USA) reportedtheir CFS precision [10] but did not provide details of theirimplementation.In our study, we develop and provide a full CFS system foraccurate measurement of the contact force between cardiactissues and catheter. In Section II, we portray what kindsof instruments were used in experiment. In Section III,we show the entire process of collection and preprocess.We collect data in environments and scenarios that arereproduced in real-world as close as possible. Through theanalysis of collected data, we identify a major problem calledShift of Reference, which hinders to accurate measurement.In Section IV, we describe the implementation details of Compressive Strain Tensile StrainShift of Wavelength of FBG under Influence of Change in Strain λλλ λ
PPPP
Reflected LightIncident Light Transmitted LightUnstrained ConditionFig. 2. A FBG sensor is a phase grating inscribed to fiber optic cable.It reflects light of certain range of wavelengths and passes that of allother ranges. Optical interrogator captures the reflected light and convertsit into digital figure. Temperature and strain are main factors that affectBragg Grating wavelength. The relationship between shift of Bragg Gratingwavelength and tension of FBG is proportional. If FBG receives compressivestrain, Bragg Grating wavelength shifts left and vice versa. our CFS system. We present two problem definitions forcontact force estimation - temporary and time series. Thecriteria for problem definition with a given experimentalenvironment are also disclosed. In line with the problemdefinition, we show that deep learning models can solveit and performance of them. Along with it, we present abaseline that anyone could follow, showing how we hadtrained. Our main contributions are the release of publiclyavailable CFS catheter dataset and being of a guideline forfuture CFS implementations.II. CONTACT FORCE SENSING SYSTEM
A. Hardware
Fig. 4 shows the design of catheter. CFS catheter wemade is divided into multiple sections. Core component istri-axial force sensor. Three were used to respond to allsituations of bending and to increase precision. They weremolded using epoxy to heat shrink tube with same distantapart. The force sensor is then placed in the central lumenof catheter to remove spatial bias. FBG sensors used formaking force sensor have (Shenzhen Lens Technology., LTD,Shenzhen, China) uniform wavelength of 1540nm, 0.5nmwidth, 10dB reflection rate and 6dB Side Mode SuppressionRatio (SMSR).Optical interrogator is another indispensable device thatsends light, measures wavelength of reflected light andconverts the wavelength to digital signal. Depending on thedevice range of wavelength, number of channels, samplingrate and precision of peak detection varies. We used FAZTI4W Interrogator (Femto Sensing International, Atlanta, GA,USA) for our experiment.Other devices are also included like a digital scale (MettlerToledo, Greifensee, Switzerland), a computer for recordingdata from interrogator and digital scale and a monitor to dis-play the results. Fig. 3 shows the CFS system composed withthe devices mentioned above. Based on the environment, wewill show how we collect FBG sensor signal data.
B. Contact Force Estimation
Using FBG sensor, we try to convert strain change in FBGto contact force of catheter - tissue contact. To do this, at first,we show that wavelength shift is related to contact force.Bragg Grating wavelength in FBG sensor is λ B = 2 n Λ where n is refractive index and Λ is periodic pitch size [4].Wavelength shift from single FBG sensor can be representedas terms of strain and temperature: ∆ λ B = 2(Λ δnδl + n δ Λ δλ )∆ l + 2(Λ δnδT + n δ Λ δT )∆ T, (1)where T is temperature and ∆ l is strain. In (1), we seethat the wavelength shift is summation of strain related termand temperature related term. At first, to consider the effectof temperature, we attach a thermocouple wire to catheter.However, because the data acquisition was done in stableroom temperature for whole time, there was no change ofb)(a) CFS CatheterInterrogatorComputer Digital ScaleFig. 3. (a) illustrates how the instruments were actually arranged during theexperiment. A computer is used for synchronized and real time acquisitionof data from both interrogator and digital scale. During the experiment,experimenter holds a catheter and pokes it onto the digital scale for 60seconds. This action is repeated with different scenarios for data diversity.(b) shows the photo of instruments used in experiments. temperature. Then we can ignore the temperature term, andrewrite (1) as ∆ λ B λ B = (1 − p e ) (cid:15) z , (2)where p e is effective strain-optic constant, (cid:15) z is axial strainand λ B is reference value of Bragg Grating wavelength.Equation (2) leaves us a linear equation between axial stran (cid:15) z and wavelength shift of FBG sensor [11].Using young’s modulus, we can obtain stress with axialstrain. By definition, we can obtain contact force using stress.So we can express contact force as term of wavelength shift:contact force = f (∆ λ B ) . (3) f can be expressed as terms of p e , strain, and etc. It istheoretically possible to predict contact force by measuringthese physical properties. However, there will be many noisesand varies by catheters or sensors. So we will find optimal f in (3) using deep learning. In other words, we will showthat young’s modulus and other physical properties can bereplaced by deep learning. (a)(b) Tip Electrode Tri-axialForce SensorElastic SpringFig. 4. (a) This diagram shows how tip section of catheter is built. Itconsists of a elastic spring, tip electrode and tri-axial force sensor built withthree FBGs. We try to built a catheter with simplest form of hardware forin hope of pursuing fundamental research. FBGs and fiber optics are notallocated in a complicated, deformed shape and unnecessary componentslike irrigation tubes are all removed. (b) This is a photo of tip section of acatheter used in experiments.
III. CONTACT FORCE SENSING DATASETAlthough CFS system is important in CAT, there is nopublic contact force sensing dataset for research. To helpwith CFS research, we construct a CFS dataset by mappingdata between interrogator and scale. To make as similaras possible to the real situation, we randomly rotate thecatheter and make it vertical to the surface. In this section,we describe how we preprocess the collected data while thequality is ensured, and some major characteristics like Shiftof Reference.
A. Peak Detection
The FBG sensor measures the intensity of each wavelengthand finds and uses wavelength corresponding to peak. Thepeak detection can be seen as important in FBG sensors. Ourinterrogator also provides peak detection function. However,the details of peak detection algorithm has not been releasedyet. By comparing to Kernel Density Estimation (KDE) peakdetection algorithm from [12], we show that interrogator’speak detection function works comparable.The idea of KDE peak detection is to make sure that eachdot is an outlier in the local context and, if correct, dismissit as a peak. Using Chebyshev inequality, we can state theconditions: Let the i th signal value be x i . Then (i) x i > m where m is the mean of entire signal and (ii) | x i − m | > h · s or some suitably chosen h > , where s is the standarddeviation [12]. KDE peak detection uses a newly calculatedvalue instead of writing the x value as it is: S ( x i ) = H w ( N ( k, i )) − H w ( N (cid:48) ( k, i )) , (4) H w ( A ) = Σ Mi =1 ( − p w ( a i ) log( p w ( a i ))) , (5) p w ( a i ) = 1 M | a i − a i + w | Σ Mj =1 K (cid:18) a i − a j | a i − a i + w | (cid:19) , (6)where N ( k, i ) is { x i − k , · · · , x i − , x i +1 , · · · , x i + k } and N (cid:48) ( k, i ) is { x i − k , · · · , x i + k } . N ( k, i ) is local context of x i without it while N (cid:48) ( k, i ) contains. KDE peak detectionuse (4), difference of entropy as a value with Chebyshevinequality. Details of the calculation of entropy can be seenin (5) and (6) For kernel function K in (6), we usedGaussian kernel: K ( x ) = √ π e − x .We test both algorithms on 0g contact force signal data.By comparing the statistic of the two algorithms, we triedto justify the Femto Sensing’s algorithm. Because we used0g contact force data, Femto Sensing’s algorithm shouldhave similar or lower standard deviation between to KDEalgorithm. As a result of the actual experiment in Table I,we were able to confirm that the two showed similar standarddeviations. It has been concluded that Femto Sensing’s peakdetection algorithm can be used as it is. B. Preprocess
Both of the collected data from interrogator and scale haveinconsistent frequency. Although ∆ t can be used together tosolve it, this makes the problem more complicate. Further-more each frequency of information has a near value. So weresample the collected data. Resampling is mainly used toestimate information in a specific time zone where informa-tion has not been collected. We use cubic interpolation as aresampling method for both data. After preprocessing, eachdata has a constant frequency of 1000 Hz and 10 Hz. C. Analysis
Our catheter’s tip is bent while collecting data and thecontact force is overly large. If two kinds of preconditionsare satisfied and large force is exerted, two FBG sensorwavelength shift positively while the other one negatively orvice versa; Fig. 2 shows the details. Two preconditions arewhen the guide wire bent the tip part of catheter or the angleof catheter-tissue is far from normal. Summing up, when theforce applied is large enough to make the bent catheter evenmore, this kind of scenario happens. Functionally, this isinevitable. If there is not enough space in the heart chamber
TABLE IS
TATISTICS OF T WO D IFFERENT P EAK D ETECTION A LGORITHM ON G C ONTACT F ORCE S IGNAL D ATA
Algorithm KDE Femto SensingSensor µ σ − − . . · sensor1 − − . . · sensor2 − − . . · ∆ λ B (nm) sensor3 Fig. 5. The above graphs are the histogram of the wavelength shift valuesof each FBG sensor. All three graphs are similar. It is a form in whichfrequency increases as it approaches zero and decreases as it moves away.Furthermore, it’s almost symmetrical so that positive and negative valuesare distributed evenly. to allow catheter to be straight, cardiologist need to bend thecatheter by one of two preconditions mentioned upper.We validate the FBG sensor data through showing theabove situation occurs evenly. In Fig. 5 which is the his-tograms of our FBG sensor data, the closer to zero, the moredata there is, and the farther away the less. This seems to bethe case because most of the time catheter is not in contact.Furthermore, all of the three FBG sensors have both positiveand negative values uniformly. Due to the randomness ofrotation angle, the uniform distribution of values means thatthe pattern mentioned above appears evenly. As a result, thedata quality is still ensured after preprocess.After training a simple Fully Connected Network (FCN),we find that there is an unexpected problem with the data.Fig. 6a shows the normal case of FBG sensor data andcontact force estimation. When the contact force reaches 0g,the wavelength of FBG sensors also returns to the referencevalue, even the FCN model also predicts 0g. Furthermore,the trend of the prediction and FBG sensor data exactlyfollows the contact force. However in Fig. 6b, we can seethat some of the FBG sensor doesn’t return to the referencewavelength. Due to this phenomenon, the FCN model startsto make a wrong prediction of contact force. In Fig. 6b,the error is almost 10g when the contact force is 0g. Wecan see that even if only one sensor went wrong, it had abig impact on contact force estimation. Furthermore, we cansee that the model’s prediction is about 0.4g floating. Thismeans that the frequency of this phenomenon is not smalland effecting the whole prediction. As a result, we decidethat this phenomenon is one of the property of our data, anddefined as
Shift of Reference (SoR) .The cause of SoR is not yet clear. Basically, the fiber a) . . . normal ∆ λ B ( n m ) time( s ) f o r ce ( g ) (b) . . . . normalSoR sensor1sensor2sensor3 time( s ) realprediction Fig. 6. The above graphs show the wavelength shift of FBG sensors according to the contact force, while the graphs below show the results of predictionsof the FCN model simply trained. (a) shows that the wavelength of the FBG sensor returns to reference wavelength when the contact force reaches 0g.Furthermore, we can see that the trend of the prediction of the FCN model follows the contact force. However, in the case of sensor 3 of (b), it can beseen that the wavelength does not return to the reference even if the contact force reaches 0g. Because the wavelength didn’t returned to reference, we cansee that the prediction is well out of 0g. We defined the phenomenon as
Shift of Reference . optics are attached to the heat shrink tube in the catheterwith either epoxy molding or UV molding. Fig. 4 showsthis structure. During the experiments, various forces areapplied like electric, thermal force. By hydraulic chamber ormechanic cable actuation, the force is transmitted over theentire body of catheter. So we expect that when experimentsbecome more frequent, the molded part is broken due tovarious forces, and the frequency of SoR is often generated.We will conduct a study to verify our expectations.IV. BENCHMARK EXPERIMENTSNot only did we collect the data, we also confirm throughbenchmark experiments that the contact force sensing actu-ally work well in our dataset. A. Problem Definition
The simplest problem definition will be predicting thecontact force for each interrogator signal. However, theinterrogator signal is approximately 1000Hz. When screengets updated with such high frequency, the readability will belargely degraded. Furthermore, the frequency of interrogatorand scale is different. For these reasons, predicting with highfrequency is inappropriate in real surgical situation. Thus,prediction rate was set to be 0.1s, meaning prediction afterlooking at one hundred interrogator signals.Let the i th interrogator signal x i . Then the t th inputof the model is X t = { x · t , · · · , x · ( t +1) − } . Usingthese expressions, the probability model predicts t th contactforce y t . Firstly, only looking at the t th input, y t can bepredicted maximizing p ( y t | X t ) . We think that looking at thewhole inputs is necessary to figure out SoR. So we defineour second probability model as time series problem - bymaximizing p ( y t | X · · · X t ) . B. Models
After we preprocess the FBG sensor data, it goes throughthe model and then the decoder layer. At last, going throughthe activation function, the output of the decoder layerbecomes the predicted contact force. The decoder layer isa linear layer that reduce the dimension to 1. We use leakyReLU for the activation function. For the intermediate model,following models are chosen.
FCN : Fully Connected Network (FCN) is made bystacking linear layers with same hidden dimension. The inputgoes through an encoder layer to fit dimensions before beingfed to stacked layers. It only looks at the current input anddoes not have any states to save the previous results.
RNN : Recurrent Neural Network (RNN) is commonlyused in time series. RNN has a state that saves previousresults. So it gives same effect as looking at the wholeinputs. There are three kinds of RNN - vanilla RNN, LSTM,and GRU. GRU [13] is chosen for the RNN cell in ourexperiment.
Transformer : Transformer is first introduced in NaturalLanguage Processing (NLP) problems [14]. It can be alsoused for time series problems by looking at the whole inputs.Furthermore our problem has constant frequency, we candirectly use the transformer. We use the decoder part oftransformer. It means that predicting t th contact force, itis not influenced by inputs after t th one. C. Training
Huber loss [15] is chosen for the loss function whiletraining, and Adam optimizer is used for the optimizerwithout any weight or learning rate decay. For learning rate,most of the models is trained by 1e-3. However, some hugemodels, such as RNN with 8 layers and Transformer with4, 8 layers, are hard to optimize with learning rate 1e-3.hese three models are trained with learning rate 1e-5 andsuccessfully converged.Table II shows the performance of predicting test datawith Mean Absolute Error (MAE) performance metric. Tofind the optimal model, we first fix the hidden dimensionand then find the optimal number of layers. All of themodel’s hidden dimension is fixed to 256. FCN model showsbest performance with 2 layers, while RNN, Transformermodels show best performance with 4 layers. After findingthe optimal number of layers, we fix the number of layers andthen vary the number of hidden dimension. For Transformer,we use 12 attention heads throughout the experiments.
D. Results
RNN model with 4 layers and hidden dimension 64achieved state-of-the-art for our dataset. For FCN model,2 layers and hidden dimension 64 showed the best per-formance. For Transformer model, 4 layers and hiddendimension 512 showed the best performance. Performanceof all 3 models did not improve because they were big.Even FCN and RNN showed the best performance whenthe hidden dimension was 64, the smallest value in ourexperiment. It showed the possibility that the problem wouldnot be so difficult and can be figure out by simple models.We expected that models which look the problem astime series will gain better performance with solving theSoR problem. RNN, one of our time series model, metour expectations. It showed a huge gap of 0.5g comparedto FCN and Transformer model. However, another timeseries model did not meet expectation. In Table II, theperformance of Transformer model is similar to FCN model.This showed that time series models do not necessarily showhigh performance. The cause of performance decline onTransformer model will be our next study.We also measured the inference time of models. Trans-former model was much bigger than FCN model with similar
TABLE IIP
ERFORMANCE ON C ONTACT F ORCE D ATASET
Model FCN RNN Transformer
NFERENCE TIME OF M ODELS ON
V100 GPU
Model FCN RNN
Average Inference time 0.759ms 1.380ms accuracy. So we only measured the inference time for FCNand RNN model. For each model, we used the optimal ones.Since catheters can be used on equipment equipped withGPUs, the inference time was checked after loading themodel on V100 GPU. Table III shows the results. Accordingto our problem definition, we have to predict contact forceat interval of 0.1s. Both models meet this requirements.V. CONCLUSIONSThrough these studies, we provide a foundation of CFSsystem. We present a full pipeline from dealing with rawFBG sensor data to contact force estimation. Furthermore,we release a contact force dataset generated by catheter thatcould serve as the basis for future research. While providingthe benchmark, we show the potential of deep learning forcontact force sensing.For future research, more techniques can be applied onour dataset. Basically, canonical filters are available in signalprocessing tasks, but they have not been applied yet [16].On the benchmark experiments, we only try fundamentalmodels: FCN, RNN, and Transformer. Since the data hasbeen validated, more complex architectures are applicablenow.For more in-depth research, there are points to supple-ment in datasets. Since the actual dataset was carried outat constant temperature during the production process, thenoise by temperature was negligible. However, temperaturechanges are frequent in RFA, cryoablation [17]. Furthermorethis dataset only considered cases where the catheter is invertical contact with the ground. In actual surgical situations,contact will be made from various angles.R
EFERENCES[1] G. Lee, P. Sanders, and J. Kalman, “Catheter ablation of atrialarrhythmias: State of the art,”
Lancet , vol. 380, pp. 1509–19, 10 2012.[2] F. Morady, “Radio-frequency ablation as treatment for cardiac arrhyth-mias,”
N Engl J Med , vol. 340, no. 7, pp. 534–544, Feb 1999.[3] R. Hamaya, S. Miyazaki, H. Taniguchi, S. Kusa, H. Nakamura,H. Hachiya, T. Kajiyama, T. Watanabe, M. Igarashi, K. Hirao, andY. Iesaka, “Management of cardiac tamponade in catheter ablationof atrial fibrillation: single-centre 15 year experience on 5222 proce-dures,”
EP Europace , vol. 20, pp. 1–7, 11 2017.[4] A. Othonos, “Fiber bragg gratings,”
Review of Scientific Instruments ,vol. 68, no. 12, pp. 4309–4341, 1997. [Online]. Available:https://doi.org/10.1063/1.1148392[5] J. Back, L. Lindenroth, K. Rhode, and H. Liu, “Three dimensionalforce estimation for steerable catheters through bi-point tracking,”
Sensors and Actuators A: Physical , vol. 279, 06 2018.[6] H. Liu, J. Back, R. Karim, K. Rhode, and K. Althoefer, “Catheter con-tact force estimation from shape detection using a real-time cosseratrod model,” 09 2015.[7] L. Sciarra, P. Golia, A. Natalizia, E. de Ruvo, S. Dottori, A. Scar`a,A. Borrelli, L. De Luca, M. Rebecchi, A. Fagagnini, A. Bandini,F. Guarracini, M. Galvani, and L. Cal`o, “Which is the best catheter toperform atrial fibrillation ablation? a comparison between standardthermocool, smarttouch, and surround flow catheters,”
Journal ofinterventional cardiac electrophysiology : an international journal ofarrhythmias and pacing , vol. 39, 02 2014.[8] F. Bourier, G. Hessling, S. Ammar-Busch, M. Kottmaier, A. Buiatti,C. Grebmer, M. Telishevska, V. Semmler, C. Lennerz, C. Schneider,C. Kolb, I. Deisenhofer, and T. Reents, “Electromagnetic contact-force sensing electrophysiological catheters: How accurate is thetechnology?”
Journal of cardiovascular electrophysiology , vol. 27, 122015.9] F. Bourier, C. Gianni, M. Dare, I. Deisenhofer, G. Hessling, T. Reents,S. Mohanty, C. Trivedi, A. Natale, and A. Al-Ahmad, “Fiberop-tic contact-force sensing electrophysiological catheters: How preciseis the technology?”
Journal of Cardiovascular Electrophysiology ,vol. 28, pp. 109–114, 09 2016.[10] S. Page and M. Dhinoja, “Smarttouch™ – the emerging role ofcontact force technology in complex catheter ablation,”
Arrhythmia& Electrophysiology Review , vol. 1, 01 2012.[11] W. W. Morey, G. Meltz, and W. H. Glenn, “Fiber Optic BraggGrating Sensors,” in
Fiber Optic and Laser Sensors VII , R. P.DePaula and E. Udd, Eds., vol. 1169, International Society for Opticsand Photonics. SPIE, 1990, pp. 98 – 107. [Online]. Available:https://doi.org/10.1117/12.963022[12] G. Palshikar, “Simple algorithms for peak detection in time-series,”01 2009.[13] K. Cho, B. van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares,H. Schwenk, and Y. Bengio, “Learning phrase representations usingrnn encoder-decoder for statistical machine translation,” 2014.[14] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N.Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” 2017.[15] P. J. Huber, “Robust estimation of a location parameter,”
Ann. Math.Statist. , vol. 35, no. 1, pp. 73–101, 03 1964. [Online]. Available:https://doi.org/10.1214/aoms/1177703732[16] S. V. Vaseghi,
Advanced Digital Signal Processing and Noise Reduc-tion . New York: John Wiley & Sons, 2008.[17] A. C. Skanes, R. Yee, A. D. Krahn, and G. J. Klein, “Cryoablation ofatrial arrhythmias,”