Atrial fibrillation (AF) poses a significant health concern due to its irregular and rapid heart rhythm. Accurate segmentation of left atrial (LA) scars and cavities from Late Gadolinium-Enhanced Magnetic Resonance Imaging (LGE-MRI) is essential for effective AF ablation procedures. In this study, we presented a novel approach utilizing a neural network architecture comprising three nnU-Nets. Departing from the original backbones, we leveraged the ResNet backbone to enhance segmentation performance. Our methodology was evaluated on the LAScarQS2022 dataset on two distinct tasks: Task-1 involved segmentation of both LA cavity and scar regions, while Task-2 focused solely on cavity segmentation. For Task-1 scar segmentation, our proposed model achieved promising results with a Dice score of 0.601. Cavity segmentation under Task-1 demonstrated even higher performance, yielding a Dice score of 0.946 and a Hausdorff distance (HD) of 10.486. In Task-2, our model achieved competitive performance metrics, with a Dice score of 0.941 and HD of 13.362. Our proposed method demonstrated superior performance, indicating its robustness and generalization capability. Our findings underscored the effectiveness of leveraging ResNet backbones and nnU-Net architecture for accurate segmentation of scars and cavities from LGE-MRI in the context of AF treatments, offering promising avenues for improved clinical outcomes in AF management.