Unsupervised phenotyping of left atrial fibrosis patterns in universal atrial coordinates from 100 LGE MRI models
K Maciunas, V Vigneswaran, A Gharaviri, M Klis, N Bodagh, A Von Kietzell, A Ranieri Guimaraes, M Williams, M O'neill, S E WilliamsAbstract
Background
Left atrial (LA) fibrosis on late gadolinium-enhanced (LGE) MRI has been linked to atrial fibrillation (AF) burden and ablation outcomes, but most studies report only global burden or manual regional scores, and there is limited coordinate-based phenotyping of LA fibrosis patterns in larger cohorts.
Objective
To identify data-driven structural phenotypes of LA fibrosis using simple region-free descriptors in universal atrial coordinates (UAC) from 100 LGE-MRI LA surface models.
Methods
We analysed 100 patient-specific LA surface meshes reconstructed from LGE-MRI with per-vertex image intensity ratio (IIR) maps and UAC parameterisation. Fibrosis was defined as IIR ≥ 1.32. To avoid unstable descriptors in almost fibrosis-free atria, we required at least 50 fibrotic vertices; 4 atria did not meet this criterion and were excluded from clustering, leaving 96 cases. For each included case we computed: (i) global fibrosis burden (percentage of fibrotic vertices); (ii) the UAC centre-of-mass of fibrotic tissue; and (iii) simple covariance-based measures of spread and orientation. These features formed a low-dimensional vector per patient. Feature vectors were standardised and clustered using k-means (k = 3). Per-cluster summaries used medians and interquartile ranges.
Results
Clustering of the 96 atria identified three structural phenotypes. A low-burden group (Cluster 3, n = 17) had very little fibrosis (median global burden ≈ 2%). An intermediate-burden group (Cluster 2, n = 51) showed modest but clearly measurable fibrosis (median ≈ 8%). A higher-burden group (Cluster 1, n = 28) exhibited greater fibrosis (median ≈ 12%) with a tail of cases exceeding 40% burden. Overall, these UAC features separated patients into low-, intermediate- and higher-fibrosis phenotypes primarily by global burden, with accompanying differences in spatial extent. Fig. 1 illustrates the distribution of global fibrosis burden across clusters, and Fig. 2 shows the UAC centres-of-mass of fibrotic tissue coloured by cluster.
Conclusion
Simple, region-free UAC-based descriptors of LGE-MRI fibrosis enabled unsupervised identification of distinct LA fibrosis phenotypes in this 100-patient cohort. These phenotypes capture both global burden and the spatial concentration of fibrosis on the atrial surface and may provide a standardised structural framework for future studies linking LA fibrosis patterns to AF mechanisms, ablation strategies and personalised electrophysiological modelling.Fig 1.Global fibrosis burden by clusterFig 2.UAC centres-of-mass of fibrotic