Left atrial (LA) segmentation is the first step in the clinical diagnosis of cardiovascular diseases, especially atrial fibrillation (AF). Therefore, there is a significant need for effective, rapid detection of LA. Unfortunately, existing methods have poor performance for cardiac details and are prone to under- or over-segmentation. To solve these problems, we report a novel LA segmentation network MFIS-net which extracts semantic features at all levels from multiple scales and uses a creatively residual module with pyramid structure to improve the fusion of deep and shallow features. Extensive experiments are conducted on the public LA challenge, the proposed MFIS-Net obtains an average Dice score of 91.6 \(\%\) and average sensitivity of 95.1 \(\%\) with less parameters, may make an important contribution to the treatment of patients with AF. All of our code will be provided upon request.