Atrial fibrillation (AF) is a prevalent heart rate arrhythmia and its incidence is increasing with the aging population. The late gadoliniumenhanced magnetic resonance imaging (LGE-MRI) provides pathologic changes in the left atrium, allowing for a detailed assessment of the left atrial anatomy, which is critical for diagnosis and treatment decisions in AF. The segmentation performance of current left atrial segmentation methods is significantly degraded when applied to multicenter data. In this work, we propose ResCAUNet, a deep learning method based on residual neural networks. Specifically, a pre-trained model driven residual segmentation network is first designed to alleviate the problem of gradient disappearance and help to extract more complex image features. Secondly, an adaptive scale weight loss function was introduced to solve the problem of discontinuous segmentation boundary, so as to ensure higher accuracy of object segmentation. Furthermore, the coordinate attention(CA) mechanism is introduced for adaptive weight allocation, focusing on the key parts of the image to improve the accuracy of left atrial reconstruction. Our method is evaluated on the LAScarQS2024 validation set and achieves an average Dice of 0.853. The evaluation results show that the proposed method is effective in left atrium segmentation of LGE-MRI.