Elevated Fibrosis Burden as Assessed by MRI Predicts Cryoballoon Ablation Failure

Abstract

Introduction:

Late-gadolinium enhancement magnetic resonance (LGE-MRI) imaging is increasingly used in management of atrial fibrillation (AFib) patients. Here, we assess the usefulness of LGE-MRI-based fibrosis quantification to predict arrhythmia recurrence in patients undergoing cryoballoon ablation. Our secondary goal was to compare two widely used fibrosis quantification methods.

Methods:

In 102 AF patients undergoing LGE-MRI and cryoballoon ablation (mean age 62 years; 64% male; 59% paroxysmal AFib), atrial fibrosis was quantified using the pixel intensity histogram (PIH) and image intensity ratio (IIR) methods. PIH segmentations were completed by a third-party provider as part of the standard of care at our hospital; IIR segmentations of the same scans were carried out in our lab using a commercially available software package. Fibrosis burdens and spatial distributions for the two methods were compared. Patients were followed prospectively for recurrent arrhythmia following ablation.

Results:

Average PIH fibrosis was 15.6±5.8% of the left atrial (LA) volume. Depending on threshold (IIR_thr), the average IIR fibrosis (% of LA wall surface area) ranged from 5.0±7.2% (IIR_thr=1.2) to 37.4±10.9% (IIR_thr=0.97). An IIR_thr of 1.03 demonstrated the greatest agreement between the methods, but spatial overlap of fibrotic areas delineated by the two methods was modest (Sorenson Dice coefficient: 0.49). 42 patients (41.2%) had recurrent arrhythmia. PIH fibrosis successfully predicted recurrence (HR 1.07; p=0.02) over a follow up period of 362±149 days; regardless of IIR_thr, IIR fibrosis did not predict recurrence.

Conclusions:

PIH-based volumetric assessment of atrial fibrosis was modestly predictive of arrhythmia recurrence following cryoballoon ablation in this cohort. IIR-based fibrosis was not predictive of recurrence for any of the IIR_thr values tested, and the overlap in designated areas of fibrosis between the PIH and IIR methods was modest.

Caution must therefore be exercised when interpreting LA fibrosis from LGE-MRI, since the values and spatial pattern are methodology-dependent.

Introduction

Atrial fibrillation (AFib) is associated with worse quality of life and increased risk of stroke and heart failure.¹ Rhythm control strategies including anti-arrhythmic drugs and catheter ablation decrease AFib burden, ameliorate quality of life, and improve cardiovascular mortality and stroke outcomes.² While catheter ablation has been shown to be superior to drug therapy in adequately suppressing the arrhythmia, its overall efficacy in eliminating AFib varies across different populations.

Atrial fibrosis measured by cardiac Late Gadolinium Enhancement Magnetic Resonance Imaging (LGE-MRI) is associated with increased rates of catheter ablation failure.³ Various studies have shown that higher LGE-MRI quantified fibrosis predicts recurrent arrhythmia following ablation, predominantly using a radiofrequency (RF) approach;^3-7 limited data are available on its predictive ability following cryoballoon ablation.⁶

Following adequate LGE-MRI acquisition, different image processing and fibrosis quantification methods can be used to predict ablation failure. One commonly-used approach uses pixel intensity histogram (PIH) of the segmented left atrial (LA) wall and delineates hyper-enhancement based on standard deviations above the mean for each patient.⁴ Alternatively, blood pool enhancement can be used as a basis for normalization, delineating regions of hyper-enhancement based on image intensity ratio (IIR) between segmented wall voxels and the LA blood pool.⁵ A comparison of these approaches to image processing and resulting quantified fibrosis and distribution has not been previous performed in a systematic and large scale fashion. LGE-MRI fibrosis quantification is poised for widespread adoption with additional applications such as computational modeling-guided ablation,⁸ and risk stratification for cardiovascular events including stroke.⁹ Thus, understanding differences between these approaches, finding common ground (a standardized cutoff threshold (i.e., IIR_thr)) for differentiating between fibrosis and non-fibrosis without treating either as a gold standard, is critical for clinical implementation of LGE-MRI, especially since different groups have reported different values.^{10, 11}

In this study, we sought to assess the usefulness of LGE-MRI-based LA fibrosis, as quantified by either PIH or IIR methods, as a means of predicting recurrent arrhythmia in a cohort of patients undergoing cryoballoon pulmonary vein isolation for the treatment of symptomatic AFib. As a secondary objective, we sought to compare fibrosis patterns produced by PIH and IIR, in terms of overall burden and spatial distribution within the LA.

Methods

Study Population:

We included 102 patients with AFib presenting for their first catheter ablation at the University of Washington using the cryoballoon catheter between July 2016 and December 2019. Patients underwent LGE-MRI imaging prior to catheter ablation as part of the standard of care. Patient demographics, ablation information and follow up post procedure were entered into the Cardiac Arrhythmia Data Repository, which is approved by the University of Washington IRB (Human Subjects Division approval #8763).

Late-Gadolinium Enhancement Magnetic Resonance Imaging:

LGE-MRI was performed up to eight weeks prior to ablation. The imaging protocol has been previously described and is summarized as follows: All studies were performed on a Philips 1.5T Achieva clinical scanner (Philips North America, Andover, MA). High resolution LGE-MRI scans of the LA were acquired 15-25 minutes following double-dose (0.2 mmol/kg) administration of gadolinium-DTPA (Multihance, Bracco Diagnostics Inc., Princeton, NJ), using a three-dimensional inversion recovery, respiration navigated, ECG-gated, gradient echo pulse sequence. Typical acquisition parameters included navigator-gated free-breathing and a transverse imaging volume with voxel size = 1.25×1.25×2.5 mm³ (reconstructed to 0.625×0.625×1.25 mm³). ECG-gating was employed to acquire a small subset of phase encoding views during atrial diastole. Cine images of the LA were used to define the interval between the R-peak on ECG and the start of image acquisition. Fat saturation sequences were used to suppress fat signal. Inversion time was identified using a scout scan of the left ventricle in short axis. Typical scan time was 5-10 minutes based on respiration and heart rate.

Image Processing and Fibrosis Quantification:

A schematic overview of the study is presented in Fig. 1. The individual methodological steps involved are described in detail in the subsections that follow. For both methods described, the amount of time required to fully process one patient’s LGE-MRI scan was ~1-2 hours in the hands of an experienced expert. The patients in this study were not part of the cohorts used to develop either method of fibrosis quantification.

Fig. 1: Visual summary of the main project elements.

NRR = non-rigid registration, which was used to deform the 3D models reconstructed from segmented IIR images to match the geometry of PIH models to facilitate quantitative comparison of fibrosis spatial patterns.

Pixel Intensity Histogram (PIH) Method.

LGE-MRI analysis to assess LA fibrosis severity was performed by the Merisight service (Marrek Inc., Salt Lake City, UT). In each patient, an ECG and respiratory-gated magnetic resonance angiogram was performed along with the 3D LGE-MRI, then both scans were sent off for segmentation and analysis. The fibrosis quantification methodology used by Merisight has been detailed in previous publications,⁶ and is briefly described here. LA wall volumes are manually calculated as the difference between epicardial and endocardial segmentations, and edited to exclude the mitral valve and pulmonary veins. The final LA segmentation includes the LA wall and pulmonary vein antra. LA fibrosis severity is quantified using a threshold-based algorithm; the threshold is selected on a slice-by-slice basis via analysis of the bi-modal intensity histogram, as described by Karim et al.¹² The proportion of voxels in the LA wall with values above threshold is calculated to derive the volumetric percentage of LA fibrosis.

Image Intensity Ratio (IIR) Method.

LGE-MRI processing was performed in-house by members of our team using the ADAS 3D software (Galgo Medical, Barcelona, Spain, and Circle Cardiovascular Imaging, Calgary, AB, Canada) to obtain a 3D reconstruction of the LA with fibrosis information. LA endocardial contours were manually drawn in the axial plane to obtain the 3D model. Further shell deformations were performed, if necessary, to ensure anatomical accuracy of the model. Fibrosis identification was based on voxel signal intensity. Boundaries of abnormal tissue were based on the previously reported normalized local IIR,¹⁰ as implemented in ADAS 3D. Briefly, the mean signal intensity of the LA blood pool was automatically calculated. Then, the IIR value for each voxel was calculated (voxel signal intensity ÷ mean voxel intensity of blood pool) and projected to the 3D model to obtain the final fibrosis map. The overall fibrosis burden using the IIR method was reported as a percentage of the LA surface area. Consistent with our goal of identifying the IIR_thr that maximizes agreement with PIH, we reconstructed models with cutoffs values ranging from 0.97 (as in Khurram et al.¹⁰) and 1.20 (as in Benito et al.¹¹), with three more values evenly distributed between these extremes (1.03, 1.08, and 1.14).

Blinding and Inter-observer variability.

Different groups of expert image processors were involved in the segmentation and analysis of LGE-MRI images for the two methods. Both groups were blinded to the results of the other method (i.e., extent and distribution of fibrotic remodeling) and to clinical outcomes of ablation procedures; notably, the IIR operators had access to the LA geometries derived as part of the PIH process, which was used to ensure similarities in macroscopic structure between the two methods. A third operator performed image registration and fibrosis comparison between the two methods and was not involved in fibrosis quantification.

Image registration and fibrosis comparison.

For the purposes of comparing the PIH and IIR methods, we derived the fibrosis burden by surface area in each LA model; this was necessary because the ADAS software does not produce an estimate of volumetric LA fibrosis burden. In each mesh, point-by-point LGE values were binarized (zero = non-fibrotic; non-zero = fibrotic, i.e., LGE+), then the fibrosis burden by surface area was calculated as the proportion of LGE+ nodes.

To facilitate point-by-point comparisons of fibrosis patterns extracted from the same MRI scans using the PIH and IIR methods, we used non-rigid registration to deform each IIR-derived mesh such that its geometry matched the PIH-derived mesh for the same patient. First, PIH and IIR meshes were converted to Wavefront .obj format and loaded into Autodesk Meshmixer (Autodesk Inc., San Rafael, CA). Then the “Align to Target” and “Attract to Target” operations were applied sequentially to deform the IIR mesh (without removing or reordering grid points) so that it registered with the PIH geometry; an example is shown in Supplemental Video 1. Finally, IIR LGE maps for different IIR_thr were transferred to the nearest nodes in the PIH mesh such that each point had two binary values (i.e., LGE± for PIH/IIR). The Sørensen–Dice coefficient (SDC) was derived in each mesh to characterize spatial overlap between PIH- and IIR-based fibrosis:

$$SDC=\frac{2\mid X\cap Y\mid }{\mid X\mid +\mid Y\mid }=\frac{2\times n_{PIH\&IIR}}{n_{PIH}+n_{IIR}}$$

where n_PIH&IIR is the number of nodes delineated as LGE+ by both PIH and IIR, while n_PIH and n_IIR are, respectively, the numbers of PIH-only and IIR-only LGE+ nodes. Notably, the IIR method includes LGE values for the PVs and the mitral valve whereas the PIH method does not; for the purposes of spatial pattern comparison and fibrosis surface area quantification, these areas were discarded from IIR meshes.

Cryoballoon Catheter Ablation:

All procedures were performed under general endotracheal anesthesia. Vascular access was obtained into the femoral veins bilaterally. Intracardiac echocardiography was used to guide trans-septal catheterization and LA access. A single trans-septal puncture was performed after heparin was administered. Ablation therapy was performed using a second generation cryoballoon catheter (Arctic Front, Medtronic Inc, Minneapolis, MN). The balloon catheter was positioned in the antral region of each of the pulmonary veins. Pulmonary vein occlusion was verified by contrast injection or intracardiac ultrasound evaluation before cryotherapy was delivered. A decapolar catheter was used to pace the right phrenic nerve while cryotherapy was delivered to the antra of the right sided pulmonary veins. A circular mapping catheter (Achieve, Medtronic Inc, Minneapolis, MN) was used to record pulmonary vein potentials. Two freezes (180 second duration) were delivered per vein. The acute procedural endpoint was electrical isolation of the pulmonary veins. According to the study protocol, no additional freezes or RF ablation were delivered outside the pulmonary vein antral regions.

Clinical Follow up and Arrhythmia Recurrence:

Patients were observed in the hospital overnight and discharged to self-care the following day. A 90-day blanking period was observed for early arrhythmia recurrences, which were addressed with cardioversions, anti-arrhythmic agents, or conservative management. Anti-arrhythmic drugs were discontinued at three months. Patients were seen in follow up at three, six, and 12 months, then every six months thereafter. Seven-day ambulatory rhythm monitors were used to evaluate for arrhythmia recurrence at each visit. 12-lead ECGs were obtained at each clinic visit and as needed based on symptoms. AFib, flutter, or atrial tachyarrhythmias lasting for longer than 30 seconds were considered recurrences according to the HRS consensus document on catheter and surgical ablation for AFib.¹³

Statistical Analysis:

Statistical analysis was performed using STATA 15 (StataCorp, College Station, TX) and Prism 9.0.1 (GraphPad, San Diego CA). We examined the distribution of all variables and summarized them using mean and standard deviation for normally distributed continuous variables; median and interquartile interval for non-normally distributed variables; or frequency and percentage for categorical variables. The association of arrhythmia recurrence and predictor variables including atrial fibrosis was examined using a survival analysis model with a Cox regression where clinical and imaging variables were studied, consistent with similar analysis in prior work.^{3, 4, 7, 10, 11}

Results

Patient characteristics.

The average age was 62±12 years and 65% of the study cohort was male. Paroxysmal AFib was present in 59.8% of patients and the prevalence of hypertension was 53.9%, whereas coronary disease and heart failure prevalence was 10.2%. Table 1 summarizes the cohort’s detailed characteristics on study entry. The mean fibrosis was 15.6±5.8% of the LA volume using the PIH method, ranging from 5.0±7.2% using the IIR_thr of 1.20 to 37.4±10.9% using the IIR_thr of 0.97.

Table 1.

Baseline characteristics of the study cohort

	N=102
Age (years)	62±12
Male sex (n;%)	66; 64.7%
Body mass index (kg/m2)	32.7±7.3
Hypertension (n;%)	55; 53.9%
Diabetes (n;%)	11; 10.5%
Coronary artery disease (n;%)	10; 9.5%
Congestive heart failure (n;%)	10; 9.5%
Prior stroke/transient ischemic attack (n;%)	3; 2.9%
Smoking history	16; 15.7%
CHA2DS2VASc score	1.9±1.4
Paroxysmal AFib (n;%)	61; 59.8%
Persistent AFib (n;%)	41; 40.2%
Prior antiarrhythmic drugs (n,%)	37; 36.3%
LA volume index (mL/m2)	74.1±29.9

Comparison of PIH and IIR geometric segmentations.

LA volumes for meshes produced by the PIH and IIR methods prior to non-rigid registration were indistinguishable, via both linear regression (Fig. 2A; R² = 0.9449; 95% confidence interval for regression line slope: 0.939–1.034) and paired analysis (Fig. 2B; Wilcoxon signed rank test; P = 0.9549). Across all models in the cohort, the median registration error between IIR and PIH segmentations was 2.3 mm (inter-quartile range: 1.57 to 3.29 mm); individual error rates for all models are shown in Fig. 2C. Registration error was larger for a handful of models (median value >4 mm in six cases), due to differences in the geometry of the segmented LA appendage. Inter-observer variability was assessed for the IIR method and agreement was high (Spearman correlation coefficient range 0.87-0.96; see Supplemental Fig. 1). Interobserver variability of the PIH method has been previously reported at 97.0%.⁷

Fig. 2: Comparison of LA geometry in 3D models reconstructed from IIR and PIH segmentations of the same images.

(A) Data plotted in x-y format with simple linear regression line, as described in text. (B) Same data in truncated violin plot, showing medians (dashed lines) and upper/lower quartiles (dotted lines); Wilcoxon nonparametric test for paired data points (p = 0.9549). (C) Box-and-whisker plots showing point-by-point registration error between IIR and PIH segmentations for each model in the whole cohort.

Comparison of PIH- and IIR-derived LA fibrosis.

Figure 3 shows examples of geometric models reconstructed via the PIH and IIR methods for several patients; for IIR methods, both pre- and post-alignment meshes are shown for different IIR_thr values as labeled. Each row also includes a map to illustrate the process used for spatial comparison of fibrosis patterns. Within each individual model, increasing the value of IIR_thr reduced fibrosis as measured by IIR (see examples corresponding to Fig. 3 cases in Supplemental Fig. 2).

Fig. 3: Comparison of fibrosis spatial patterns in models reconstructed from PIH and IIR image segmentations.

Columns 2 and 3 show the IIR-derived pattern before and after non-rigid registration (NRR). Regions of spatial overlap are shown in column 4; corresponding Sørensen-Dice Coefficient (SDC) values are shown. In rows A-C, different values of IIR_thr are used to delineate fibrosis in IIR-based models. See also maps showing IIR-derived models for all IIR_thr values for these cases in Supplemental Fig. 2.

In terms of quantification of fibrosis burden (Fig. 4A), the best match between surface area PIH and IIR was for IIR_thr = 1.03 (PIH: 19.9% [13.7%; 26.0%] vs.18.8% [13.9%; 24.8%]; p=0.69). Of the five IIR_thr values tested, this was the only one for which there was no significant difference between fibrosis burdens derived by the two approaches. In terms of spatial overlap (Fig. 4B), the best match between PIH and IIR was also found for IIR_thr = 1.03 (SDC: 0.46 [0.37; 0.51]); notably, the difference between SDC values for this threshold and those associated with either IIR_thr = 0.97 or 1.08 were not statistically significant. When the analysis conditions were relaxed to allow for the best IIR_thr value to be used in each patient (rather than choosing one IIR_thr to be used across the board), the spatial overlap improved (Fig. 4C; global SDC across all five groups: 0.49 [0.40; 0.55]). As shown in Fig. 4D, there was no apparent pattern linking the overall extent of fibrotic remodeling and the tendency towards particular optimal IIR_thr values; similar plots of fibrosis burden from IIR vs. PIH for all individual thresholds are shown in Supplemental Fig. 3.

Fig. 4: Quantitative comparison of PIH- vs. IIR-based fibrosis across all models.

(A) Fibrosis surface area for PIH models (column 1) vs. IIR models with different IIR_thr values; n = 102 in all columns; Dunn’s multiple comparisons test P<0.0001 for all pairs except as labeled. (B) Sørensen-Dice Coefficient (SDC) for quantification of spatial overlap between fibrosis patterns between PIH models and IIR models with IIR_thr values as shown; n = 102 in all columns; Dunn’s multiple comparisons test P<0.0001 for all pairs except as labeled. (C) Patient-by-patient plot showing SDC values for the optimal IIR_thr value only for all 102 models. (D) x-y plot showing the relationship between the IIR fibrosis in the optimal-IIR_thr model vs. PIH fibrosis for the same individual; same color-coding as in (C). See similar x-y plots for each individual IIR_thr value in Supplemental Fig. 3.

Arrhythmia recurrence and Predictors.

Atrial arrhythmias occurred in 42 patients (41.2%) outside an initial blanking period of 90 days. The average follow-up period was 362±149 days and the median time to recurrence was 208 days. Recurrence rates were higher among persistent AFib patients compared to paroxysmal (53.7% vs 32.8%; p=0.036).

Univariate analysis of predictors of arrhythmia recurrence demonstrated that persistent AFib was associated with higher risk of AFib recurrence (HR 1.99, p= 0.028). PIH-derived fibrosis was also associated with AFib recurrence (HR = 1.08, p = 0.019). IIR-derived fibrosis did not predict arrhythmia recurrence, regardless of IIR_thr value.

Multivariate analysis conducted including 4 variables including persistent AFib, LA volume, LA fibrosis determined by PIH, and congestive heart failure, selected based on statistical significance in univariate analysis, demonstrated that LA fibrosis determined by the PIH method independently predicted AFib recurrence with a HR of 1.07; p=0.036 (Table 2). When fibrosis was grouped into the previously described Utah staging groups, arrhythmia recurrence occurred in 5 of 18 (21.7%) patients with stage 1 (<10%) fibrosis, increasing to 23 of 59 (39.0%) in patients with stage 2 (10-20%) fibrosis) and 13 of 19 (68.4%) in patients with stage 3 (20-30%) fibrosis. This relationship is demonstrated by Kaplan Meier curves in Fig. 5. Receiver-operator curve analysis was also used as a means of predicting AFib recurrence via fibrosis burden (see Supplemental Fig. 4). The area under the curve (AUC) for binary prediction of AFib recurrence via PIH fibrosis was 0.600, with sensitivity and specificity values of 0.75 and 0.47, respectively, at the optimal fibrosis burden cut-off of 12.6%. These metrics of classification performance were generally superior to those seen for the IIR method (AUC range: 0.490 to 0.535; sensitivity range: 0.47 to 0.75; specificity range: 0.38 to 0.56).

Table 2.

Univariate and multivariate survival analysis results using a Cox regression, reporting hazard ratios for recurrent arrhythmia.

	Univariate Analysis			Multi-variate analysis
	Hazard Ratio	95% CI	P value	Hazard Ratio	95% CI	P value
Age	0.99	0.97-1.02	0.659
Female sex (vs male)	1.44	0.72-2.88	0.304
Body mass index (kg/m2)	1.00	0.95-1.05	0.926
Hypertension	1.29	0.69-2.42	0.421
Diabetes	1.26	0.50-3.23	0.624
Coronary artery disease	1.82	0.76-4.33	0.177
Congestive heart failure	1.87	0.83-4.23	0.134	2.05	0.86-4.90	0.107
Prior stroke or transient ischemic attack	2.498	0.60-10.39	0.208
Smoking history	0.65	0.26-1.67	0.374
Persistent AFib (vs paroxysmal)	1.99	1.08-3.69	0.028	1.36	0.67-2.82	0.393
Prior antiarrhythmic drugs	1.22	0.65-2.29	0.531
LA volume index (mL/m2)	1.01	0.99-1.02	0.054	1.00	0.99-1.01	0.763
PIH fibrosis (vol. %)	1.08	1.01-1.11	0.019	1.07	1.01-1.12	0.036
IIR fibrosis (area %) 0.97 threshold	2.58	0.12-53.20	0.540
IIR fibrosis (area %) 1.00 threshold	2.43	0.11-52.92	0.572
IIR fibrosis (area %) 1.03 threshold	2.69	0.11-64.57	0.542
IIR fibrosis (area %) 1.05 threshold	2.97	0.11-77.29	0.512
IIR fibrosis (area %) 1.08 threshold	3.74	0.12-111.94	0.447
IIR fibrosis (area %) 1.14 threshold	7.68	0.18-333.35	0.289
IIR fibrosis (area %) 1.20 threshold	18.72	0.27-1290.47	0.175

Fig. 5: Kaplan Meier survival curves demonstrating time to recurrent atrial arrhythmia in three groups of atrial fibrosis quantified using the PIH method and grouped by previously published stages.

Only one patient had PIH fibrosis >30% and did experience recurrence; this individual is not shown. The log-rank test of equality of survivor functions between stages demonstrated a statistically significant difference in arrhythmia recurrence between the groups (p=0.01).

Discussion

In this study, we examined differences in LGE-MRI quantified atrial fibrosis using two widely available methods in a cohort of 102 patients who underwent their first catheter ablation procedure using the cryoballoon. Our main findings are: (1) PIH-based volumetric assessment of atrial fibrosis was modestly predictive of arrhythmia recurrence following cryoballoon ablation in this cohort; (2) IIR-based fibrosis was not predictive of recurrence for any of the IIR_thr values tested, which spanned the range of cutoffs suggested by prior studies;^{10, 11} (3) the agreement in overall quantified fibrosis and overlap in designated areas of fibrosis between the PIH and IIR methods is modest; and, (4) an IIR_thr of 1.03 is associated with the best agreement between the two methods. Atrial fibrosis is a microscopic phenomenon observed in many cardiovascular diseases, including AFib, and consists of myocyte loss, collagen deposition, and extracellular space expansion.⁷ These structural changes lead to reduced electrical conduction velocity and wavefront block, giving rise to an AFib substrate.¹⁴ Cardiac MRI’s spatial resolution is orders of magnitude too coarse to resolve such microscopic details, but LGE-MRI leverages the kinetics of the extracellular contrast agent gadolinium to provide us with a comprehensive, non-invasive assessment of the tissue-scale consequences of fibrotic remodeling. Knowledge of atrial fibrosis degree can be used to guide personalized care of patients with clinical and pre-clinical AFib.^{9, 15} Prior studies have shown how fibrosis level can be used to identify patients who are not likely to respond to ablation therapy for AFib. The ALICIA¹⁶ and the DECAAF-II^{17, 18} clinical trials, which used the IIR and PIH methods respectively, evaluated the strategy of targeting MRI identified fibrotic tissue compared to standard of pulmonary vein isolation; neither study reported improvement in arrhythmia recurrence with the fibrosis-tailored approach. In addition, fibrosis based computational models of AFib inducibility have been demonstrated to identify ablation targets resulting in better ablation outcomes.⁸ The OPTIMA clinical trial (NCT04101539), currently enrolling, is a randomized analysis to assess the efficacy of PVI compared to a computationally guided ablation, previously shown to be effective in a pilot study,⁸ in patients with persistent AFib and evidence of LA fibrosis. Moreover, LGE-MRI quantified fibrosis has been associated with embolic stroke of undetermined source in the absence of AFib,¹⁵ and computational modeling has suggested that these patients may have latent potentially pro-arrhythmic substrate.¹⁹

Successful LGE-MRI atrial fibrosis quantification is a two-step process, with (A) adequate image acquisition followed by (B) image processing for identification of hyper-enhancement for fibrosis quantification. Both methods of fibrosis quantification studied in our paper have been shown to identify fibrosis and predict ablation failure.^4-7 To the best of our knowledge, this is the first study to examine these two methods applied to the identical MRI scans from the same cohort of ablation-naïve patients. Prior studies using IIR have reported fibrosis quantified using different IIR_thr values.^{10, 11} By definition, increased IIR_thr reduces the extent of tissue delineated as fibrotic, reducing the sensitivity with which AFib substrate is identified. High IIR_thr values (≥1.14) generally resulted in poor agreement with PIH in terms of both overall burden and distribution of fibrosis. On the opposite end of the IIR_thr spectrum, the lowest threshold (0.97) led to increased sensitivity and overestimation of the fibrotic burden compared PIH. IIR_thr = 1.03 was the operating point at which the two methods had the greatest agreement in overall fibrosis quantified. For quantitative assessment of the agreement in fibrosis spatial patterns, an SDC value of 1 would indicate complete overlap of the fibrotic areas identified by PIH and IIR methods. Under the analysis condition when all 102 models were assessed using the same IIR_thr value, the best SDC values (0.46 [0.37; 0.51]) were observed for IIR_thr = 1.03; when we optimized for the best IIR_thr per patient, SDC values improved (0.49 [0.40; 0.55]). Interestingly, within the latter group the modal IIR_thr value was 0.97 (i.e., the optimal IIR_thr value was 0.97 for more cases than any other IIR_thr value including 1.03), suggesting a general trend towards lower IIR_thr values resulting in better agreement with PIH models in this data set. Overall, the modest level of spatial overlap between the two methods (~0.5 at best) indicates that there are areas of identified fibrosis exclusive to only one method or the other that are at least equal to areas of fibrosis identified by both. These discrepant areas largely consist of PIH-identified fibrosis at higher IIR_thr and IIR-identified fibrosis at lower IIR_thr. Moreover, the IIR method selects a smaller sample of the LA wall (mid wall) and thereby fewer pixels to identify fibrosis compared to a larger sample (endo- and epi-segmentations) for the PIH method. Of note, the correlation between the LA volumes obtained by the two methods was excellent (and inter-rater agreement between two different expert users examining the same LGE-MRI scans in ADAS 3D), indicating that differences in the distribution of fibrosis were not due to discrepancies in the volumetric rendering of the LA.

The results presented here are somewhat unexpected in the broader context of AFib research using LGE-MRI for fibrosis quantification. In our hands, the highest IIR_thr value tested (1.20) suggested remarkably low fibrosis across 102 patients (1.98% [0.94%; 5.05%]). Other groups have used even higher IIR_thr values and observed much higher fibrosis levels (e.g., 25.1% [16.8%; 31.4%] with IIR_thr = 1.22).⁸ Since the IIR approach itself is straightforward, it seems unlikely that this discrepancy is due to image processing errors but rather likely due to differences in the cohort studied, raw imaging data output by various MRI scanners at different hospitals, and/or different gadolinium-based contrast agents used. This points to the importance of future research in this area, including multi-site harmonization to provide consistent imaging studies to compare the essential properties of LA LGE-MRI scans performed at different centers on distinct imaging hardware.

Our study is not the first to attempt comparison of different schemes for quantifying LA fibrosis assessed by LGE-MRI. Notably, in a study that predated the development of the IIR approach, Karim et al.¹² established a “pseudo-ground truth” by combining manual segmentations of LA scar, then compared the performance of 11 candidate approaches for threshold-based segmentation. The SDC results in that study resembled those seen in this work, with all median values <0.5 and some individual SDCs as low as ~0.05. In a more recent study, Hopman et al.²⁰ systematically compared IIR to a different automatic method based on intensity normalization (voxels ≥n standard deviations above the blood pool mean delineated as fibrotic) in a study of 47 patients who underwent LGE-MRI prior to AF ablation. There were major discrepancies in the extent of LA fibrosis when thresholds previously identified as optimal (IIR_thr = 1.2 and 3SD, respectively) were used; notably, their results agree with ours in terms of very high and very low fibrosis burdens for IIR_thr values on the lower and higher end of the published range (0.97 and 1.20). In a more recent study focused on a cohort of 37 ablation-naïve patients, Eichenlaub et al.²¹ concluded that there were large discrepancies between the extent of LA fibrosis estimated by the IIR and PIH approaches, which is highly consistent with our own findings. Lastly, data presented in a preprint by Nairn et al.²² assess differences in LGE-MRI delineation approaches in the context of intracardiac voltage mapping and characterization of left atrial conduction velocity. Their findings confirm that important discordances exist in the extent and spatial localization of identified pathological LA substrate depending on the LGE-MRI segmentation method used. The totality of this research, conducted by different groups at different centers in different patient cohorts, is quite sobering. Since clinical outcomes of ablation were not discussed in the latter studies, the authors could not definitively identify a preferred approach between IIR, PIH, or 3SD. Alongside the present results and given the potential discrepancies in images acquisition discussed in the preceding paragraph, these studies highlight potentially major deficiencies in methods based on normalization to blood pool intensity. Overall, this reinforces the need for more work in this area, up to and including multi-center comparison studies that examine discrepancies between fibrosis quantification approaches, especially in the context of clinical risk stratification.

The PIH-identified volumetric fibrosis estimate emerged in multivariate modeling to be significantly associated with recurrent AFib in this cohort; since the software used for IIR segmentation does not compute a volumetric fibrosis estimate, a head-to-head comparison between the two approaches was impossible. Notably, neither IIR-based nor PIH-based estimates of LA fibrosis surface area were associated with recurrent AFib.

Prior studies have largely included RF ablation patients, with a small number of patients undergoing cryoballoon ablation.⁶ All ablations in this cohort targeted pulmonary vein isolation with no extra-pulmonary vein targets ablated using the cryoballoon or additional RF ablation. The overall recurrence rate of this mixed paroxysmal and persistent AFib cohort was 42% after about one year, which is comparable to reported rates from clinical trials and cohort studies using cryoballoon ablation. The hazard ratio associated with recurrent AFib with PIH fibrosis was 1.07 indicating that a 1% increase in fibrosis was associated with a 7% increase in the risk of recurrence. This is consistent with findings from the DECAAF study, in which most patients underwent RF ablation.⁶ Our study thus reaffirms the utility of fibrosis quantification in patient selection for ablation where patients with severe fibrosis are unlikely to benefit from cryoablation and suggests that pre-ablation fibrosis is an unmodifiable risk factor for recurrence in the context of this type of procedure. There is no reason to believe the discrepancy between IIR and PIH quantification systems is due to the decision to include only cryoballoon ablation patients. In fact, the decision to study these patients was motivated by the fact that in cryoballoon procedures, tissue outside the pulmonary veins is seldom targeted due to the device design; in contrast, wide area circumferential RF ablation is subject to inter-operator variability.¹³

The AUC values we report for fibrosis-based binary prediction of AFib recurrence are modest (0.600 for PIH), but this must be in interpreted in the context of the near-complete absence of stage 4 (>30%) fibrosis patients in our cohort. If more patients with extreme fibrosis had been included, it is reasonable to expect the classification performance of this analysis would be improved; however, in the context of studies like DECAAF, such patients are seldom referred for catheter ablation at our center based on the expectation of poor clinical outcomes.

Limitations.

This study was conducted in a single center, but the sample size, imaging, and image processing expertise were robust to perform the analyses described. The PIH method fibrosis results were not performed on site; rather, they were used as reported to our center through a clinical service agreement by Merisight, Inc. We therefore did not have the ability to study various threshold results using this method. The IIR software was provided to us by ADAS 3D and was used as marketed clinically by Circle Inc. With these limitations, the comparison of the two image processing methods was done without bias to the results of the other method and with blinding of the image processing team to arrhythmia recurrence. Nevertheless, we acknowledge that the “black box” nature of Merisight’s third-party segmentation and fibrosis quantification approach is scientifically frustrating, particularly in a comparison study like this. However, from our standpoint this shortcoming is significantly offset by the fact that this system is integrated into the day-to-day workflow at our hospital as standard of care for patients undergoing AF ablation, which makes it very attractive and convenient as a research platform. Our intention was to study the methods as they are currently implemented in clinical practice. The IIR approach as implemented in ADAS 3D allows users to select an intensity threshold, while the PIH approach as carried out by Merisight does not. As such, we believe our decision to treat PIH as a reference and IIR as a comparator in the present study was justified.

Since IIR requires voxel normalization to the mean LA blood pool intensity, it is susceptible to imaging artifacts or acquisition defects including poor nulling of the left ventricular myocardium. This may explain why IIR-based quantification produced very high estimates of fibrosis burden in some cases (e.g., outliers with >60% fibrosis for all IIR_thr values, as in the upper part of Fig. 4A). In an IIR-only study it may have made sense to censor these scans; however, in the present study, we opted not to exclude any potential patients, since we wanted to maintain the largest sample size possible for the AF recurrence part of the study and because the PIH approach was evidently not susceptible to the same artifacts or defects.

Fibrosis patterns for the two segmentation approaches were compared point-by-point following use of a non-rigid registration process to transform each IIR-based model into the same geometry as the PIH-based model (see Supplemental Video 1). It is possible the registration process introduced error that influenced the apparent mismatch in fibrosis patterns. However, our calculations suggest that the overall registration error was relatively small (~2 to 3 mm). Given the raw MRI resolution and the fact that fibrotic feature sizes in these reconstructions are generally on the order of >1-2 cm, we argue that it is unlikely the point-by-point comparison of fibrosis patterns between PIH and IIR segmentations are markedly affected by registration error.

In addition to pre-ablation fibrosis, it is possible that the quality of lesions created during ablation might also affect outcome. Since post-ablation MRIs were not part of the standard protocol for this cohort, we were unable to assess that relationship in this study. In future research it may be possible to correlate the quality and extent of cryoablation lesions with AF recurrence. Considering this study’s findings, interactions between pre-ablation fibrosis and lesions created during treatment might be a potential avenue for investigation.

Clinical and translational outlook.

Fibrosis is a central piece of the atrial disease puzzle, which includes AFib and other adverse outcomes of stroke and heart failure. Current research is bringing us closer to understanding processes that determine the development and progression (or perhaps regression?) of fibrosis. Systematic understanding of LGE-MRI-based quantification of fibrosis and acknowledgement of its current limitations and areas for improvement are crucially important to accompany advances in fibrosis research. As applications of fibrosis quantification in arrhythmia, stroke, and heart failure management continue to evolve, the cardiology community, particularly cardiac electrophysiologists and imaging experts, must shoulder an outsized responsibility in ensuring that non-invasive imaging methods accurately represent tissue-level biological phenomena.

Conclusions.

In this large cohort of 102 patients who underwent cryoballoon AF ablation with lesions limited to the pulmonary vein region, excessive fibrosis quantified by PIH was associated with failure, but the same was not true when the IIR approach was used, even when a wide range of thresholds was tested. As such, caution must be exercised when interpreting LGE-MRI, since fibrosis burdens and spatial pattern of remodeled tissue are methodology-dependent. More work is needed in this area to arrive at the definitive approach for LA fibrosis quantification, up to and including a multi-center study carefully comparing the many possible methodologies and ensuring the results are properly analyzed in the context of meaningful clinical outcomes.

Abbreviations:

AFib

Atrial fibrillation

PIH

Pixel Intensity Histogram

IIR

Image Intensity Ratio

LGE-MRI

Late-gadolinium enhancement Magnetic Resonance Imaging

Data availability:

Patient-derived data are not publicly available due to privacy concerns. Parties wishing to obtain these data for non-commercial reuse should contact the co-corresponding authors.