Publications

Arachnoid Membrane Segmentation in Intraoperative Microscopic MVD Surgery Scenes

​

MICCAI Workshop 2025 (COLAS)

 

Abstract: 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 highquality, 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 membraneâĂŹs 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 decisionmaking.

Intraoperative Absolute Depth Estimation in MVD Surgery

​CBMS 2025 

Abstract: Microvascular decompression (MVD) is a neurosurgical procedure that relieves nerve compression by repositioning or separating offending blood vessels, effectively reducing pain or spasms. Accurate localization of the compression site is crucial for optimal surgical outcomes, as it enables precise identification and decompression of the offending vessel. While horizontal anatomical relationships are easily identified in the surgical view, compressions occurring along the depth axis are more challenging to discern. In this study, we propose a method to measure accurate intraoperative distances during MVD surgery using DepthAnything-V2. By leveraging the optical properties of standard imaging equipment in conjunction with the depth estimation model, our method computes precise, absolute distances rather than relying solely on relative measurements, achieving distance estimation errors of less than 2 mm compared to intraoperative and preoperative reference measurements.

Developing the surgeon-machine interface: using a novel instance-segmentation framework for intraoperative landmark labelling

​

Abstract: Artificial intelligence (AI) has the potential to enhance intraoperative safety, surgical training, and patient outcomes by enabling real-time interpretation of complex surgical environments. In this study, we introduce the concept of the Surgeon–Machine Interface (SMI) to describe the interaction between surgeons and machine inference, and present a custom deep computer vision framework designed to operationalize this concept under a sparse labeling paradigm. Using a modified SOLOv2-based architecture, our platform performs precise instance-level segmentation of anatomical landmarks and surgical instruments from intraoperative microscopic video of open spinal dural arteriovenous fistula (dAVF) surgery, despite being trained on a limited number of annotated frames. The proposed SMI outperformed state-of-the-art segmentation models, including Mask R-CNN, YOLOv3, and SOLOv2, achieving higher F1-scores, mean Average Precision, and faster inference times, while demonstrating strong generalization to previously unseen objects and out-of-domain scenarios. These findings highlight the promise of AI-enabled SMI platforms for high-efficiency intraoperative landmark guidance and establish a foundation for future real-time neurosurgical guidance systems scalable across surgical procedures.