Introduction: Atrial fibrillation (AF) treatment is often impeded by limited anatomical insights. Late Gadolinium-Enhanced MRI (LGE-MRI) has emerged as a valuable tool for predicting the success of ablation procedures. In this study, we present a deep learning approach for enhanced segmenting atrial chambers and walls using 3D LGE-MRI, developed as part of our participation in the MICCAI Multi-class Bi-Atrial Segmentation (MBAS) 2024 Challenge. Methods: We employed a preprocessing pipeline to standardize intensity values and enhance image contrast. Multiple state-of-the-art models were utilized for segmentation: both 2D and 3D nnU-Net architectures, Swin-UNETR, MedNeXt, and U-Mamba. To optimize segmentation performance, we experiment with various loss functions, including Dice Loss, TopK Loss, Boundary Loss, and Hausdorff Distance Loss. Results: In the validation phase of the challenge, our approach achieved the second-place ranking. Utilizing a 3D nnU-Net architecture trained with a Dice Loss function, we obtained Dice Scores of \(66.55\%\) for the Atrial Wall, \(88.25\%\) for the Right Atrium, and \(91.14\%\) for the Left Atrium. Conclusion: The MedNext-B model achieved the best performance on our internal validation set using a combined loss function incorporating DC, HD95, TopK, CE, and Boundary losses. However, the difference in performance compared to using only DC and CE losses was not statistically significant. To meet the inference time requirements, we used data cropping.