Objective: Quantitative analysis of spindle dynamics in mitosis through fluorescence microscopy requires tracking spindles elongation in noisy image sequences. Deterministic methods, which use typical microtubule detection and tracking methods, perform poorly in the case of the sophisticated background of spindles. Methods: In this paper, we present SpindlesTracker, a fully automatic and extensible workflow that can efficiently analyze time-lapse images’ dynamic spindle mechanism. First, we design a novel deep neural network: YOLOX-SP (YOLOX for spindle). It consists of double branches responsible for spindle bounding boxes and endpoints detection. Then an improved SORT algorithm is used to link the same identity in different frames. Subsequently, we pair endpoints that fall into the same spindle bounding box as the spindle poles. Finally, we introduce the minimal cost path (MCP) algorithm to extract the continuous, single-pixel spindle skeleton. Result: SpindlesTracker is evaluated in all aspects of detection, tracking, and skeleton extraction through a fission yeast dataset. It achieves 84.1% mAP in bounding box detection and over 90% accuracy in endpoint detection. And for tracking, the comparison results show that the improved SORT algorithm increases by 1.3% in multiple object tracking accuracy (MOTA) and by 6.5% in multiple object tracking precision (MOTP). In addition, the statistical result shows that the mean error of spindle length is within 1 ?m. Conclude: SpindlesTracker provides a new baseline for multiple spindles analysis. Significance: This workflow could be easily extended to other microtubule or filamentous structures. The code is released on GitHub.