U-10Zr-based nuclear fuel is pursued as a primary candidate for next-generation sodium-cooled fast reactors. However, more advanced characterization and analysis are needed to form a fundamental understating of the fuel performance, and make U-10Zr fuel qualify for commercial use. The movement of lanthanides across the fuel section from the hot fuel center to the cool cladding surface is one of the key factors to affect fuel performance. In the advanced annular U-10Zr fuel, the lanthanides present as fission gas bubbles. Due to a lack of annotated data, existing literature utilized a multiple-threshold method to separate the bubbles and calculate bubble statistics on an annular fuel. However, the multiple-threshold method cannot achieve robust performance on images with different qualities and contrasts, and cannot distinguish different bubbles. This paper proposes a hybrid framework for efficient bubble segmentation. We develop a bubble annotation tool and generate the first fission gas bubble dataset with more than 3000 bubbles from 24 images. A multi-task deep learning network integrating U-Net and ResNet is designed to accomplish instance-level bubble segmentation. Combining the segmentation results and image processing step achieves the best recall ratio of more than 90% with very limited annotated data. Our model shows outstanding improvement by comparing the previously proposed thresholding method. The proposed method has promising to generate a more accurate quantitative analysis of fission gas bubbles on U-10Zr annular fuels. The results will contribute to identifying the bubbles with lanthanides and finally build the relationship between the thermal gradation and lanthanides movements of U-10Zr annular fuels. Mover, the deep learning model is applicable to other similar material micro-structure segmentation tasks.