Brain tumor surgery presents major challenges due to the complex anatomy of the brain and the infiltrative nature of these lesions, which are often located near eloquent areas. One of the key determinants of patient survival is the extent of tumor resection, provided it can be achieved without compromising neurological function 1. To maximize safe resection, neurosurgeons rely on a variety of intraoperative adjuncts, including fluorescent agents, neuronavigation, direct electrical stimulation, and advanced intraoperative imaging techniques, most notably intraoperative magnetic resonance imaging (ioMRI) and intraoperative ultrasound (ioUS) 2.
Although ioMRI offers excellent resolution and accuracy, its high cost, logistical demands, and complexity of integration limit its availability to a small number of specialized centers 3. In contrast, ioUS is a low-cost, versatile modality that integrates naturally into the surgical workflow 4-6. Nevertheless, its broader adoption has been limited by several factors: high operator dependency, a steep learning curve, and interpretation challenges related to artifacts, non-standard imaging planes, low contrast between tumor and normal brain, and a restricted field of view.
Over the past decades, research on AI-based segmentation of brain tumors has advanced substantially, but most work has focused on MRI 7. In the context of ioUS, early studies such as Ritschel et al. 8 demonstrated that supervised classification models (e.g., support vector machines) could distinguish tumor from healthy tissue in contrast-enhanced ultrasound, but these approaches were labor-intensive and limited to small datasets. Ilunga-Mbuyamba et al. 9 later investigated multimodal registration between ioUS and MRI to enhance segmentation, but clinical feasibility was constrained by the need for accurate co-registration. More recently, deep learning-based approaches by Canalini et al. 10 and Carton et al. 11 have been applied to segment surgical cavities and tumor volumes in ioUS images.
State-of-the-art methods such as those reported by Faanes et al. 12, using nnU-Net architectures, have achieved promising Dice similarity coefficients of 0.6-0.9 on public datasets such as RESECT-SEG 13 and ReMIND 14. However, these models were not designed for real-time inference and have not undergone validation in live surgical settings. Other approaches, such as that of Dorent et al. 15, have relied on synthetic ultrasound images derived from preoperative MRI, raising concerns about generalizability to real ioUS data. Overall, despite these advances, clinical translation remains limited due to the unique challenges of ioUS, including lower spatial resolution, image heterogeneity, and variability in acquisition protocols.
In other medical domains, AI-assisted ultrasound segmentation has demonstrated real-time feasibility. For example, Hu et al. 16 implemented U-Net-based models for breast lesion segmentation at 16 frames per second (FPS) with Dice scores exceeding 0.75, while Wei et al. 17 applied YOLO-based detection to identify carotid plaques with 98.5% accuracy at 39 FPS. Despite their efficiency and accuracy, similar approaches have yet to be implemented and clinically validated for brain tumor surgery using ioUS.
Our project aims to address this gap by conducting a multicenter, prospective, non-randomized validation of a prototype deep learning-based segmentation model specifically designed for real-time intraoperative brain tumor ultrasound. The model operates at surgical frame rates, automatically delineating tumor boundaries directly on the live ultrasound feed, with the goal of assisting intraoperative decision-making, maximizing the extent of resection when oncologically appropriate, and preserving neurological function.
This study builds upon our prior work, which has established a strong scientific foundation for the proposed validation. In our recent publication, "Deep Learning-Based Glioma Segmentation of 2D Intraoperative Ultrasound Images: A Multicenter Study Using the Brain Tumor Intraoperative Ultrasound Database (BraTioUS)" 18, we trained and validated a deep learning model for brain tumor segmentation using multicenter data from BraTioUS-DB-marking a milestone in ioUS research. In a second study, "Real-Time Brain Tumor Detection in Intraoperative Ultrasound: From Model Training to Deployment in the Operating Room"19, we developed and prospectively evaluated a real-time computer vision detection model in the operating room. Together, these contributions provide a robust framework for advancing real-time AI-based segmentation in intraoperative ultrasound, directly aligned with the objectives of the present study.
