Accurate segmentation of bi-atrial structures and their walls in patients with atrial fibrillation is essential for detailed anatomical analysis and patient-specific treatment planning. Determining the thickness of the atrial wall is particularly challenging due to significant regional variability and low image resolution, which existing approaches often overlook by applying a fixed thickness. Accurate segmentation of both atria and their walls can substantially improve outcome of catheter ablation by providing patient-specific treatment strategies. To address challenges in atrial segmentation, we developed and validated Segmentation of Bi-Atria and Wall Network (SegBAW-Net), a multistage deep neural network designed to automatically segment the left and right atria along with their walls. The performance of SegBAW-Net was evaluated using the Dice Similarity Coefficient and the 95% Hausdorff Distance metrics. Data triaging was applied to ensure robust training and validation. The network was trained using 3D late gadolinium-enhanced magnetic resonance images provided by the MICCAI MBAS 2024 Challenge, which included 70 scans for training, 30 for validation, and 100 for testing. A key contribution of this paper is the in-depth analysis of both the data and ground truth by two experts.