Accurate coronary artery segmentation is essential for tasks such as stenosis detection, anatomical labeling, intervention planning, and computational modeling in coronary artery disease (CAD). However, coronary segmentation remains challenging due to issues such as vessel overlap, noise, and severe class imbalance. This study introduces Centerline Focal Loss (FLOw-Loss), a novel loss function, designed to improve coronary segmentation in 2D X-ray coronary angiography (XCA). FLOw-Loss combines the structural sensitivity of centerline Dice (cl-Dice), which focuses on centerline alignment, with the class-balancing strengths of focal loss. This enables FLOw-Loss to capture fine vessel details more effectively and improve topological consistency. Our approach employs the nnU-Net framework, optimized with FLOw-Loss, and trained on the ARCADE dataset to perform binary vessel segmentation. It is externally validated on three additional open datasets (DCA1, XCAD, and FSCAD). Evaluation metrics include Dice, cl-Dice, recall and false negative rate (FNR), as well as masked variants (mDice and mcl-Dice) to handle partially annotated data. Results show that FLOw-Loss outperforms standard and topology-aware baselines across most metrics, especially in masked centerline overlap and vessel detection. Our approach enhances the robustness and continuity of coronary segmentation under varied clinical conditions and annotation standards, making it a valuable tool for CAD analysis in real-world XCA data.