Timing of acquisition
The finding that aCNR and scar detection increased with time is not surprising. In an informative study by Goldfarb et al. [
28], the T
1 values for LV myocardial scar, viable myocardium and blood pool at 2 min intervals following GBCA administration up to 1 h. Theyfound that the discrimination between scar and viable myocardium was significant even at very early acquisitions (< 10 min). However, the discrimination between blood pool and scar was only significant at > 10 min, and continued to improve with time, such that imaging at > 30 min was recommended for blood pool to scar differentiation.
For PAAS imaging, it is the blood pool to scar differentiation that is crucial, not viable myocardium to scar. All centers currently acquire 3D LGE imaging with in-plane resolution around 1.3 × 1.3 mm (Table
1), and therefore blood pool partial voluming effects are inevitable for most voxels within an atrial wall of thickness 2-4 mm [
29,
30]. PAAS detection will be improved as the blood pool signal falls, regardless of the image interrogation technique. This is most critical when the maximum intensity projection technique is used to interrogate scar [
3,
8‐
10], but the principle also applies for a voxel-by-voxel interrogation of the atrial wall [
2,
7,
11,
31].
In addition, it should be noted that the ‘true’ threshold for atrial scar, if it is even appropriate to binarise scar and healthy atrial myocardium at all, almost certainly changes with timing of acquisition. 3.3 SD from the blood pool mean was selected as an objective threshold with histological validation [
13], but those pre-clinical scans were only performed at approximately 20 min post GBCA administration, and the confidence intervals were wide. Likewise, an equivalent threshold of IIR 1.32 at 20 min was also selected in view of demonstrated clinical correlation, but the derivation of the threshold was from ventricular myocardial imaging [
24].
Clearly, there will be a tight correlation between aCNR and %PAAS at a fixed threshold, and the increase in %PAAS scar detected with time should be interpreted with caution. At earlier acquisitions, the ‘true’ threshold, representing an optimal compromise of sensitivity and specificity, is likely to be lower as the scar signal intensity lies closer to that of the blood pool. Instead, increasingly high thresholds may be best viewed as an index of the confidence with which portions of the atrial shell may be classified as neither blood pool nor healthy myocardium. The innate advantage of the maximum intensity projection technique and the ‘z-score’-type threshold is that it provides a degree of quantification of the likelihood that a signal intensity on the shell could possibly be derived from a blood pool voxel. Further work is required in order to define the appropriate thresholds for delineation of PAAS at different timings post GBCA, but this study has suggested that there is a decrease in the level of noise in scar detection with time from GBCA administration. The lower panels of Fig.
8 show that areas that have not been ablated may be inappropriately delineated as scar at early acquisitions. Later acquisitions, using a higher threshold, are less likely to detect coincidental noise thus increasing confidence in scar (higher specificity for PAAS detection at later acquisitions). The improvement in DSC, reflecting appropriate scar detection, from 20 min to 30 min was small but significant.
The DSC values obtained in this study are lower than those found in a benchmark study comparing post-ablation scar locations to expert-derived pseudo-truth (0.72–0.85 for algorithm-derived scar, 0.14–0.59 for fixed thresholds) [
32]. However, it is important to consider that there are two additional factors that will contribute to a DSC lower than those achieved for analysis of PAAS on the same raw image. Firstly, for this study PAAS was compared to ablation on a different LA anatomical shape, acquired on EAM. There will inevitably be regions of mis-registration of the relatively narrow bands of scar. Secondly, scar location is clearly not perfectly recorded on EAM. An average location of energy delivery is recorded, but is confounded by contact force, stability, and other determinants of scar formation. Therefore, many locations where energy is delivered will not have true scar formation. An average DSC of 0.39 is therefore in the anticipated range for a good assessment of ablation scar.
The timing of 3D LGE acquisition varies widely in published studies, with no center routinely imaging at > 30 min (Table
1), although it should be noted that some vendors such as Siemens Healthineers routinely acquire central k-space in the middle, rather than start, of the scan, effectively pushing the scan acquisition timing back by half of the scan duration compared to this study. There have been two large non-selective studies of PAAS imaging, where CMR imaging was performed regardless of recurrence status. Both acquired 3D LGE sequences at around 15 min post-GBCA administration and gaps in ablation lines were frequently detected. Badger and co-workers detected gaps in PV scar at 405/576 veins (70%) and in 93% of patients overall [
3]. Akoum and co-workers, on assessment of a subset of the DECAAF study, detected circumferential scar at 1.26 veins per patient (gaps estimated at 67% of veins) and also in 93% of patients overall [
2]. The findings of the present study suggest that the incidence of gaps may hypothetically have been substantially lower if image acquisition had been delayed until later after GBCA administration.
Scan parameters
The improvement in PAAS imaging with reduced, single, GBCA dose relates largely to the increase in blood-pool-to-scar-contrast, and the superiority is most marked at early acquisitions. The persistence of the improved aCNR at 20 and 30 min is an interesting illustration of the aphorism that ‘less is more’ and is not necessarily apparent from first principles: this is an important finding of this study. The relationship between contrast concentration and signal intensity is not a linear one, and the halving of contrast concentration in any given compartment does not necessarily result in a halving of relaxation rate (the inverse of the relaxation time constant). Similarly, the increase in signal resulting from a shortened T
1 time-constant is a relationship that is also highly dependent upon inversion time and repetition time [
33]. Furthermore, the time dependent concentrations of GBCA within the blood pool and atrial myocardial scar compartments have not been clearly ascertained [
33].
The lack of improvement in imaging at 3 T may be explained at least in part by challenges in image acquisition that are more frequently encountered in this environment. ECG interference is higher, leading to triggering errors, and the respiratory navigator is less reliable, although once successfully commenced overall acquisition time was unchanged from controls. Contrast behavior is also relatively unchanged, with minimal reduction in relaxivity of GBCAs at higher field strengths (5.0 mmol
− 1 s
− 1, (range 4.7–5.3 mmol
− 1 s
− 1) at 3 T in plasma, versus 5.2 mmol
− 1 s
− 1 (range 4.9–5.5 mmol
− 1 s
− 1) at 1.5 T) [
34]. However, the acquisition window was late atrial diastole (onset 296 ± 40 msec post R-wave), which was more frequently impinged upon at the longer inversion times necessary for imaging at 3 T, requiring compromise in terms of acquisition window. The mean inversion times at 10 min, 20 min and 30 min were 238 msec, 267 msec and 288 msec respectively, and the acquisition window had to be delayed for one subject at 10 min, two at 20 min and four at 30 min. Generally the impingement upon the window was only by 10-20 msec, but the maximum impingement for any subject was 80 msec, markedly increasing the nominal acquisition time. Finally, many of the acquisition parameters clearly cannot be directly transposed from the 1.5 T to 3 T environment, and in particular compromises regarding receiver bandwidth, TE and TR had to be made, which may also have impacted upon imaging quality.
There was a general decline in imaging quality with half-slice thickness, which is not surprising. The reduced voxel size will decrease the voxel SNR, but on direct image assessment the blood pool and scar aSNR remained relatively preserved, as was aCNR (Additional file
2: Figure S1 and Additional file
3: Figure S2). However, when defining scar tissue at 3.3SD above the blood pool mean there was a significant decrease in PAAS area overall. This has implications for the detection of small gaps. In the recent study by Bisbal et al., they found median gap size of 13 mm, but the smallest was 1.6 mm [
7], and Ranjan et al. detected deliberate gaps as small as 1.4 mm, using a 1.0 × 1.0 × 1.5 mm resolution 3D LGE acquisition in an animal model [
14]. Small gaps will only be detectable within plane for thicker slice 3D acquisitions, and not if the gap lies between slices. Two consecutive orthogonal acquisitions may represent the best compromise for accurate gap detection whilst maintaining scar sensitivity, but would require more complex registration and co-processing for gap detection.
Clinical implications
PAAS imaging in the immediate term presents opportunities for non-invasive evaluation of conventional and novel therapies. This includes assessment of the impact of contact force [
35], evaluation of ablation extent by cryoballoon [
36], and even ablation-induced modification of fat pads containing ganglionated plexi [
37]. Optimal, and ideally uniform, imaging acquisition parameters would increase precision and facilitate comparison of studies.
The use of PAAS imaging to guide ablation procedures is more controversial. Interscan reproducibility needs to be demonstrated, and sensitivity needs to improve. However, if the findings of Bisbal et al. can be replicated then there is opportunity for swifter and more efficacious re-do procedures [
7]. This may become even more relevant in the light of the PRESSURE trial where it was demonstrated that there was increased arrhythmia free survival with prophylactic repeat ablation procedure at 2 months, regardless of recurrence status [
38]. Non-invasive CMR correlates that identify subjects who would benefit from pre-emptive repeat procedures could be extremely valuable.
For post-ablation macro re-entry arrhythmias, identification of PAAS may also assist in the pre-procedural prediction of the arrhythmia mechanism. This in turn may inform activation mapping strategy, diagnostic manoeuvres and possibly lesion delivery. Zahid et al. used LA LGE datasets to derive patient specific models of LA tachycardia pathways, in combination with fibre orientation atlas. In 7 out of 10 patients (all post–PVI) it was possible to model a LA macro-reentrant circuit, and the ablation trajectory that was successful clinically was predicted in-silico in all 7 patients [
39].
Limitations
This study was performed at 3 months post ablation, using Gadovist as the GBCA, and is an evaluation of chronic scar formation. As such, the results are not directly applicable to the assessment of acute lesion formation, and could not be used to guide acute repeat ablation during the index procedure in a hybrid-type environment. Likewise, there is evidence that there is a slow fading of scar with time [
40], and the application of these results to imaging > 3 months post-ablation, or using different contrast agents, should be performed with caution.
There is no gold standard for validation of PAAS detection, in the absence of histological assessment. Manual segmentation was considered, but strongly relies upon subjective user-defined thresholding of the scar and was therefore rejected. Voltage mapping has been only weakly correlated with PAAS, and it is likely that registration errors, bipolar sampling considerations and electrode size confer upon voltage mapping a similar level of error as CMR assessment of scar. Furthermore, there is evidence that voltage and true scar are only moderately well-correlated [
19]. Therefore, the study has focused on optimising sensitivity, rather than evaluations of specificity of scar detection.
In terms of the study design and the analysis of the impact of the timing of the acquisition post-gadolinium, it is important to note that the study is underpowered to detect small differences with time from GBCA administration for the non-standard acquisition parameters. In addition, the acquisitions could not be performed at identical timepoints post-GBCA administration, which may introduce a bias for late acquisition for patients that experienced more difficult and prolonged imaging acquisitions. On account of technical considerations, it was not possible to randomize patients to the 3 T scanner, and additionally scanning parameters cannot be replicated exactly between different scanner field strengths. The 3 T acquisition parameters aimed for equivalence, rather than optimisation for 3 T imaging alone, and the results at the higher field strength should be interpreted with caution. The interval between scan sessions was minimized in order to control for time dependent scar maturation [
40], but there was a possibility of residual GBCA accumulation between scans. T
1 relaxation times for the myocardium were unchanged between scan sessions, and there was no systematic difference between scans in any parameter for control patients. Recent studies have suggested that very low concentrations may persist beyond 48 h [
41], despite the interval being > 20 half-lives, but the impact on the results is likely to be minimal.