Accurate detection and segmentation of brain metastases (BM) are essential for stereotactic radiosurgery (SRS) planning. This study first systematically compares six loss functions across three representative 3D deep learning models. Results show Dice is a robust baseline, CE improves precision, and Focal has limited effectiveness. While JVSS substantially enhances small-lesion sensitivity, it reduces precision; combined losses achieve the most balanced and robust performance. Furthermore, we propose a novel inference strategy: local overlap fusion for subvolume merging (LOF_SM). By applying axis-wise local weighting in overlapping regions, LOF_SM significantly improves computational efficiency, reducing inference time by approximately 30%, while simultaneously maintaining sensitivity and improving precision over conventional sliding window (SW) methods. These findings underscore the importance of jointly optimizing loss functions and inference strategies to achieve a superior balance between detection performance and computational efficiency, thereby facilitating the clinical application of BM segmentation.