Microvascular decompression (MVD) is a neurosurgical procedure to treat cranial nerve compression syndromes such as trigeminal neuralgia and hemifacial spasm. The arachnoid membrane (AM) is a thin, transparent meningeal layer that adheres to or covers neurovascular structures and must be carefully dissected to access the surgical site during MVD surgery. Proper AM dissection is essential for visualizing the operative field and ensuring safe vessel and nerve manipulation. However, AM dissection is technically challenging due to its poor contrast with surrounding tissues and close adherence to critical neurovascular structures. To address this, we propose the first dedicated study on AM segmentation from operative MVD videos. We introduce a high-quality, expert-annotated dataset focusing on AM in the cerebellopontine angle and train a segmentation model with a task-specific loss function to improve AM segmentation. Our results demonstrate that the proposed loss function improves AM segmentation performance by 7.35% in IoU over the baseline, enabling reliable segmentation despite the membrane’s transparency and intraoperative variability. This work lays the foundation for automated AM recognition in surgical environments and provides a valuable resource for AM dissection and surgical decision-making.