Measuring Lesion Gaps, Not Scar Volume, to Predict Recurrence After Atrial Ablation from Small Cohorts

Catheter ablation treats atrial fibrillation (an irregular, chaotic heart rhythm) by burning or freezing lines of scar that electrically wall off the tissue where the arrhythmia starts. Twenty to fifty percent of patients relapse within a year (Predicting and Preventing Post-Ablation Arrhythmia Recurrence). Invasive remapping shows that relapse often follows a small break in one of those lines. In the Marshall-PLAN cohort, one incomplete line lowered 12-month success from 72% to 48%, and 22% of roof lines thought to be blocked had hidden gaps (Takagi et al.). Current image-based models ignore this. Segmenters are scored on volume overlap, so a 1 mm break in a 40 mm line costs almost nothing. Recurrence predictors are end-to-end networks that need hundreds of patients, which most ablation centers do not have.

We propose to compute a small number of physically meaningful "gap scores" from post-ablation contrast MRI. Late gadolinium enhancement (LGE) MRI is a scan in which scar lights up. Each score is the width of the widest healthy-tissue channel crossing an intended lesion line, measured along the curved atrial wall rather than through 3D space, and averaged over segmentation uncertainty. We then feed these few scores to a time-to-event survival model. Because the representation is fixed by conduction physics rather than learned, it should predict recurrence from fewer than 100 patients. It should also predict the type of recurrence: pulmonary-vein gaps should lead to fibrillation and isthmus gaps to organized flutter-like tachycardia. A black-box model cannot make that prediction.

Core Claim

Much of the recurrence signal in post-ablation imaging comes from whether each lesion line is continuous, not from how much scar there is. A gap score defined on the atrial surface should therefore add prognostic value beyond scar volume and clinical covariates. That added value should hold at small sample sizes, where learned image models lose to simple baselines. The claim can be falsified in three ways. The gap score may add nothing over scar percentage in a nested survival model. It may fail to predict simulated inducibility when gaps are known exactly in silico. Or gap location may fail to predict recurrence phenotype.

Research Program

Uncertain scar maps on the atrial surface. We segment the left atrial wall and scar with a joint model in the style of JAS-GAN, which reports scar Dice of 0.821. We project the result onto a triangulated wall mesh. We keep an ensemble of segmentations, not a single mask, because the gaps we care about are near the resolution limit of LGE MRI. To register intended lesion corridors across patients, we use a statistical shape model. These corridors are the pulmonary vein antra, the roof, and the mitral, dome and cavotricuspid isthmuses. We build the shape model with the PCA shape-space approach of Bhalodia et al. and the pulmonary-vein-branching shape priors of Sodergren et al.. Those papers use shape as the predictor; we use shape only as a coordinate system for locating lines.

Gap score. For corridor $c$ on surface $M$, let $d_s(x)$ be the geodesic distance from surface point $x$ to the nearest scar. Geodesic distance is measured along the wall, not through the blood pool. The healthy-channel width is

$$w_c = \max_{\gamma \in \Gamma_c} \; \min_{x \in \gamma} \; 2\,d_s(x),$$

where $\Gamma_c$ is the set of surface paths that cross the corridor from one side to the other. This is a widest-path problem, which Dijkstra-type algorithms solve on the mesh. The patient-level feature is $p_c = \Pr(w_c > w^*)$, taken over the segmentation ensemble, where $w^*$ is a conduction threshold calibrated in simulation. Measuring along the surface borrows the argument of the multi-fidelity Gaussian-process paper: its Laplace–Beltrami kernels showed that Euclidean distance leaks information across walls that are not connected. We apply that reasoning to lesion continuity rather than to mapping inducibility.

In-silico calibration where gaps are known. Zolotarev et al. predicted acute termination of fibrillation in 1,000 virtual hearts, but every virtual lesion was assumed complete. We take the same kind of synthetic anatomies and the pacing-simulation setup of the multi-fidelity study. Into the ablation masks we insert gaps of controlled width and location, then simulate inducibility. This gives the threshold $w^*$. It also tests whether $p_c$, computed from degraded, MRI-like versions of the masks, recovers the true simulated outcome. This stage also fills an acknowledged gap in the digital-twin literature.

Clinical survival modeling under scarcity. The corridor scores, around five to seven numbers per patient, enter a penalized Cox model alongside scar percentage and a few clinical covariates. The time-to-event outcome is first documented recurrence lasting 30 seconds or more, with right-censoring. We also fit a competing-risks model for the two phenotypes, fibrillation versus organized tachycardia. When a larger source cohort with overlapping variables is available, we add instance-reweighted transfer in the style of Transfer Survival Boosting. That method's reported gain was on tabular heart-failure data and has never been tested in ablation. Transfer is an optional layer, not the core of the proposal.

Data. Post-ablation LGE MRI with scar labels exists in the style of the 192-scan JAS-GAN cohort and the automated scar workflows of Malcolme‐Lawes et al.. The file lists no public cohort that pairs post-ablation LGE with follow-up outcomes. The clinical arm therefore needs a partner center with 60 to 150 patients who have post-procedure MRI and rhythm follow-up. Ideally a subset would also have invasive remapping at a redo procedure, which gives ground-truth gap locations.

Baselines

The naive reference is the CHA₂DS₂-VASc score, a standard bedside stroke-risk tally. Qiu et al. report an AUROC of 0.538 for it, and a clinical-only model reached 0.642. We will refit both on our cohort, since those numbers come from a different population.

The most direct competitor uses the same scan summarized differently: a Cox model on total scar volume or scar percentage, computed from the same segmentation ensemble. The only difference is how the scar map is summarized. JAS-GAN shows scar volume can be measured reliably (r = 0.974–0.987 against manual estimates). No reported result was found on Lacuna for scar volume as a predictor of recurrence. To beat this baseline, the gap scores must significantly improve the likelihood ratio and C-index when added to the scar-volume model, not merely match it. Here C-index is the probability that the model ranks two patients' recurrence times in the right order.

The strongest learned competitors are two end-to-end models. The first is a 3D CNN trained on the LGE images with shape-space augmentation, following Bhalodia et al. The second is the ECG-plus-clinical dual-branch Transformer of Qiu et al. That Transformer reports AUROC 0.899 at a training size of 644. No reported result was found on Lacuna for it below 100 patients. We will run both under matched conditions: identical patients, repeated nested cross-validation, a fixed hyperparameter-search budget of equal GPU-hours, and learning curves at 25, 50, 100 and the full cohort. Beating them means a higher cross-validated C-index at 100 patients or fewer. At full size we will report honestly whether they catch up.

Two ablations isolate the key ingredient. The first replaces geodesic widths with Euclidean widths. The second replaces the ensemble probability $p_c$ with a single-mask binary gap indicator. For the in-silico stage, the reference point is Zolotarev et al.'s reported ROC-AUC of 0.92 for full multimodal inputs and 0.69 for the ablation mask alone. Both were obtained on lesions assumed complete. We will re-evaluate a model of that type on the gap-injected cohort.

Why not just…?

Why not just use scar volume? This is the most serious threat. Patients with less total scar may simply have more gaps, and scar percentage could capture most of the signal. We do not argue this away. The nested Cox comparison against scar volume is the headline test, and a null result there would falsify the core claim. The in-silico stage separates the two quantities by construction: we inject gaps while holding scar volume fixed. The Marshall-PLAN finding of 72% versus 48% success depending on line completeness is the reason to expect continuity to matter separately from volume.

Why not just train a CNN end-to-end on the post-ablation MRI? Bhalodia et al. show that end-to-end prediction from MRI works when heavy augmentation is available. Their augmentation deforms anatomy, though, and never creates or removes a millimeter-scale break in a scar line. At fewer than 100 patients, the review file's evidence (Chest X-ray study; Bria et al.) warns that high-capacity models pick up site artifacts instead. The learning-curve comparison will show where, if anywhere, the CNN overtakes the fixed features.

Why not just fine-tune the ECG–clinical Transformer from Qiu et al.? It is the best reported clinical model, but it cannot say which line failed. An ECG recorded in sinus rhythm also reflects the atrial substrate more than the ablation lines. We include it as a baseline and also test the combination of ECG embedding plus gap scores. If the combination beats both inputs alone, the two sources carry complementary information.

Why not just add a topology-aware loss to the segmenter? That would make masks more continuous, but it is the wrong fix. Near the resolution limit the key question is whether a gap exists at all, and a loss that enforces continuity would bias the answer toward "no gap." We keep segmentation standard and represent uncertainty explicitly through the ensemble. The single-mask ablation shows how much that uncertainty contributes.

Decisive Checks

We will run four decisive tests. The first is in silico: at fixed scar volume, $p_c$ should predict simulated inducibility, while a Zolotarev-style mask model's AUC should drop from its complete-lesion level. The second is clinical: adding gap scores to scar percentage plus clinical covariates should give a significant likelihood-ratio improvement and a higher cross-validated C-index at 50 and 100 patients. The third is phenotype specificity: pulmonary-vein-corridor scores should predict fibrillation relapse, and isthmus scores should predict organized tachycardia, echoing the 67%/60% split in Takagi et al. The fourth uses the remapping subset: corridor scores should localize the invasively confirmed gaps better than chance. Passing the first two supports the claim. Passing the third and fourth shows the scores reflect the mechanism that the black-box models cannot identify.