4243

Toward automatic lesion transmurality assessment using machine learning: a proof of concept in preclinical EP studies under MRI-guidance
Valéry Ozenne1,2,3,4, Pierre Bour2,3,4, Marylène Delcey2,3,4, Nicolas Cedilnik5, Maxime Sermesant5, and Bruno Quesson2,3,4
1Centre de Résonance Magnétique des Systèmes Biologiques, UMR 5536, CNRS, Bordeaux, France, 2IHU Liryc, Electrophysiology and Heart Modeling Institute, Fondation Bordeaux Université, Bordeaux, France, 3Univ. Bordeaux, Centre de recherche Cardio-Thoracique de Bordeaux, U1045, Bordeaux, France, 4INSERM, Centre de recherche Cardio-Thoracique de Bordeaux, U1045, Bordeaux, France, 5Université Côte d’Azur, Inria, Epione, Sophia Antipolis, France

Synopsis

MR-guidance of electrophysiological (EP) procedures requires manual segmentation of the cardiac cavities either at the beginning of the procedure to produce the roadmap volume or after radiofrequency ablation (RFA) to assess the lesion transmurality in post ablation images. The purpose of this work is to evaluate the feasibility of automatic in-line segmentation in the context of routine preclinical EP studies.


Introduction

There is a substantial need for improving electrophysiology (EP) ablation treatment in complex arrhythmia like ventricular tachycardia [1,2,3]. Indeed, fluoroscopy-guided cardiac catheterization lack of visualization of the lesion formation during the procedure. The adoption of EP studies under MRI guidance includes the possibility of immediate visualization in response to catheter ablation using MR thermometry and post-ablation visualization using long-T1 inversion recovery sequence (Fig.1). Such workflow requires manual segmentation of the cardiac cavities either at the beginning of the procedure to produce the roadmap volume or after radiofrequency ablation (RFA) to assess the lesion transmurality in post ablation images.

Methods

Acquisition: CMR was performed on a 1.5 T MR system using two clinically used 16-channel cardiac coils. At the beginning of the procedure, a 3D bSSFP (or roadmap) sequence was performed in order to visualize the full anatomy and to segment the ventricular cavities. After one or multiple RFA, a 3D navigator-gated Turbo Flash IR pulse (or post-ablation) sequence was performed with long inversion time (TI) for blood suppression in order to visualize the lesion without the use of contrast agents (Fig.2). Preprocessing: retrospective bias correction using the N4 algorithm [4] was applied on all images followed by an adaptive, patch-based denoising algorithm[5]. Template-based database creation: In order to ease the segmentation process, two representative templates for roadmap and post-ablation images were constructed from the respective population using ANTS [6] (Fig.3). Manual segmentation of the template segmentation were mapped on each training data by application of the inverse transformations. Dual 3D U-nets: localization - cropping and segmenting: a successive dual 3D U-Net architecture [7] was used (Fig.4). The first network localized the heart position. The second networks performed individual segmentation of the ventricular cavities. The U-net model used for training independently 34 roadmap volumes and 24 post-ablation volumes. Testing were done for each model on 4 additional volumes coming from further experiments. Lesion transmurality assessment: The wall thickness map of the left ventricle were computed from the epicardium and endocardium masks. Then, an arbitrary threshold was applied on 3D long-TI sequence on the left ventricle to extract the mask of the lesions. The latter was finally multiplied by wall thickness map to obtain the prediction of the lesion transmurality. Automatic pipeline was designed using the Gadgetron framework [8] and BART [9] (Fig.5).

Results

The training time was 14 hours for the localization step and took approximately the same time for the individual segmentation steps. Weighted dices loss were stable or slowly decreasing after iteration 150 (Fig.6). Accuracy for the roadmap volumes (mean +/- standard deviation) was 0.92+-0.02 (LV endo) 0.86+0.01 (LV epi) 0.87+-0.05 (RV endo) 0.64 +-0.11 (RV epi). Model has been used in further experiments: Segmentation of the cavities took approximately 3 min and were used in EP procedure for catheter navigation in the interventional EP software (Fig.8). Prediction of the lesion transmurality was also evaluated but after the experiment (Fig.9).

Conclusion

Automatic in-line ventricular segmentation during preclinical interventional CMR EP procedure is feasible. The data availability limits the accuracy of the method as well as the presence of catheter artefacts in black and hyper signal in the epicardial fat and the created lesion. Image integration of MR information could help the cardiologist to precisely quantify the size and the lesion transmurality during the procedure.

Acknowledgements

This work was supported by the "Agence Nationale de la recherche" under the program ‘‘Future Investments’’ with the reference ANR‐17‐CE19‐0007 (CARTLOVE), ANR-10-IAHU-04 (IHU LIRYC).

We also would like to thanks Tom Lloyd and Jason Stroup from Imricor Medical Systems that provide the MR compatible ablation catheters used in this studies

References

Tanner H et al. Catheter ablation of recurrent scar-related ventricular tachycardia using electroanatomical mapping and irrigated ablation technology: results of the prospective multicenter Euro-VT-study. J Cardiovasc Electrophysiol. 2010;21(1):47-53.

Berte B et al. Impact of septal radiofrequency ventricular tachycardia ablation: insights from magnetic resonance imaging. Circulation. 2014;130(8):716-8.

Wittkampf FH et al. RF catheter ablation: Lessons on lesions. Pacing Clin Electrophysiol. 2006;29(11):1285-97. doi: 10.1111/j.1540-8159.2006.00533.x. PubMed PMID: 17100685.

Tustison NJ et al. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging. 2010 Jun;29(6):1310-20. doi: 10.1109/TMI.2010.2046908. Epub 2010 Apr 8. PMID: 20378467; PMCID: PMC3071855.

Manjón J et al. Adaptive non-local means denoising of MR images with spatially varying noise levels. J Magn Reson Imaging. 2010 Jan;31(1):192-203. doi: 10.1002/jmri.22003. PMID: 20027588.

Avants BB, Tustison NJ, Stauffer M, Song G, Wu B, Gee JC. The Insight ToolKit image registration framework. Front Neuroinform. 2014 Apr 28;8:44. doi: 10.3389/fninf.2014.00044. PMID: 24817849; PMCID: PMC4009425.

Shuman Jia et al.. Automatically Segmenting the Left Atrium from Cardiac Images Using Successive 3D U-Nets and a Contour Loss. STACOM: Atrial Segmentation and LV Quantification Challenges, Sep 2018, Granada, Spain. pp.221-229,

Hansen, M. S., & Sørensen, T. S. (2013). Gadgetron: an open source framework for medical image reconstruction. Magnetic resonance in medicine, 69(6), 1768-1776

Uecker, M., Tamir, J. I., Ong, F., & Lustig, M. (2016). The BART Toolbox for Computational Magnetic Resonance Imaging. ISMRM.

Figures

Figure 1: Schematic view of the EP workflow under MRI guidance. A 3D bSSFP acquisition is used as a reference for the roadmap & EP mapping. The cavities are segmented and converted to meshes. Then catheter navigation is performed using active catheter tracking followed by EP mapping. RFA ablation is performed under MRI temperature monitoring followed by post-ablation imaging without contrast agent. Evaluation of the lesion size and transmurality is done on the post ablation images after epi/endo segmentation. The segmentation steps could benefit from machine learning.

Figure 2: Specific features of the 3D images. Top: 3D bSSFP acquisition or 3D roadmap acquisition after catheter introduction into the cavities. The artefacts of the catheter (red arrows) are larger than the actual size of the catheter, in particular at the vicinity of the tip. The epicardial fat in hyper signal( blue arrows). Bottom: 3D long-TI acquisition performed after RFA. The created lesion can be visualized in hyper signal (yellow arrows) compared to the surrounding tissue attributed to edema (green arrows). The epicardial fat is highly visible in anterior and posterior view.

Figure 3: Database creation. For both datasets, bias correction using the N4 algorithm and patch-based denoising algorithm were first applied on all images. Then representative templates from respective population were constructed using ANTS. Manual segmentation of the cavities on the template images were performed and finally mapped on each training data by application of the inverse transformations used in the template construction.

Figure 4:Proposed segmentation workflow. The 3D whole heart volume were first cropped in a unique volume around the heart cavities using a CNN model. Then individual CNN models were used to segment the endocardium and epicardium of each ventricular cavities. The resulting segmentation were then merged into a single label map.

Figure 5: Proposed fully automatic in-line segmentation workflow: In-house reconstruction was performed in-line on an additional workstation using the Gadgetron framework. After coil-compression, parallel imaging and compress sensing reconstruction were performed on the GPU using the Berkeley Advanced Reconstruction ToolBox inside the Matlab gadget. In line prediction of the contour cavities was predicted on the fly in a python Gadget including Keras and Tensor flow. Reconstructed images were automatically send back to the MR console.

Figure 6: Weighted dice loss of the training and validating cases. 3D bSSFP volume (top) and 3D long TI (bottom) volume weighted dice loss are showed for each cavities and each contours. In both cavities, validation loss of the endocardium contours performed better than for the epicardium contour.

Figure 7: The dice overlap coefficient for the 3D bSSFP sequence and the 3D long TI sequence. As mentioned in the method part, the data were split per experiments to build independent training and test data. As a consequence and due to the limited number of available datasets, the test datasetwas built with four volumes coming from two different experiments for the 3D bSSFP and with four volumes coming from only one experiment for the 3D long TI which results in a small standard deviation for the latter.

Figure 8: Automatically segmented cavities on a test case. Segmentation of the endocardium and epicardium are shown on the Monte Carlo platform (Siemens Healthcare) during EP studies. An additional step was needed to convert the masks into meshes for display.

Figure 9: Comparison of transmurality assessment. Top view. Template based segmentation (ground truth) of the left and right epi/endo in 3D rendering and 2D orthogonal views. Wall thickness and lesion transmurality were computed from the endo/epi masks. The colormap indicates the percentage of wall transmurality over the total wall thickness (blue 10%, green 50%, red dark 90%). Bottom view. Segmentations predicted from the individual CNN model. The result were similar except in the basal anterolateral part where epicardial fat was include in the LV mask and then detected as a lesion.

Proc. Intl. Soc. Mag. Reson. Med. 29 (2021)
4243