Introduction: Left atrial (LA) fibrosis and epicardial adipose tissue (EAT) are spatially linked to the arrhythmogenic substrate in atrial fibrillation (AF) and can be quantified using late gadolinium enhancement (LGE)-MRI and CT. Accurate quantification may improve understanding of AF pathophysiology and support risk stratification. However, prior studies have reported inconsistent findings due to heterogeneous imaging protocols, differing definitions of LGE and EAT, and reliance on manual segmentation.

Goal: This thesis aimed to develop a standardized, reproducible method to quantify LA fibrosis and LA-EAT on LGE-MRI and CT in AF patients before and after ablation. The method will be applied to multimodality data from patients undergoing catheter ablation at Pitié Salpêtrière Hospital in Paris.

Methods: A deep learning-based pipeline was developed using publicly available datasets, combining two nnU-Net models to crop around the LA cavity and then segment the LA cavity and LA wall. LGE was quantified using multiple threshold-based methods (n-SD, n-BP, IIR) and validated against expert-annotated scar masks. EAT was quantified by dilating the LA cavity mask, obtained with two 3D U-Nets, and extracting voxels in the –200 to –50 HU range from contrast-enhanced CT scans. The final pipeline was applied to LGE-MRI and CT images from patients undergoing catheter ablation for AF. Associations between quantified substrate were evaluated.

Results: The nnU-Net achieved Dice scores of 0.933±0.023 (LA cavity) and 0.654±0.078 (LA wall) on the MICCAI 2018 LGE dataset. The optimal scar quantification method consisted of the mean blood pool intensity plus 2.3 times its standard deviation, achieving a Dice score of 0.395 on 60 annotated LA volumes from the LAScarQS 2022 challenge. The method was applied to 74 pre-ablation LGE-MRI volumes and 73 patient-matched CT scans, yielding a median %LA-LGE of 13.0 [6.3 - 22.5]% and LA-EAT index (LA-EATI) of 4.1 [3.3 - 4.6] mL/m². %LA-LGE was significantly correlated with normalized LA volume (r: 0.382, p: 0.001) but not with LA-EATI. Volume comparison between the LA cavity predicted on LGE-MRI versus ground truth mask obtained on CT showed high correlation (r: 0.889, p<0.001). In 16 post-ablation scans, %LA-LGE was significantly higher post-ablation (13.5 [6.5 - 21.0] vs 23.2 [14.7 - 27.6] %, p = 0.008), while a non-significant decrease in LA volume was observed (56.8 [50.6 - 71.8] vs 53.3 [49.3 - 63.7] mL/m², p:0.065). Visual inspection showed increased enhancement near pulmonary veins in cases with large %LA-LGE change.

Conclusion: This automated method enables standardized quantification of LA-LGE and LA-EAT in AF patients and successfully detected post-ablation changes. While performance was affected by segmentation reliability, which was influenced by image quality among other factors, the findings support its clinical relevance. Further improvements could enhance robustness and enable broader applicability.