Intraoperative ultrasound-based brain tumor segmentation with boundary-overlay visualization for surgical assistance

Brain shift during neurosurgical tumor resection reduces the reliability of preoperative image-based navigation and increases the need for updated intraoperative image interpretation. Intraoperative ultrasound (iUS) provides accessible intraoperative anatomical information but remains difficult to interpret because of speckle noise, weak contrast, and ambiguous tumor appearance. This study aims to develop an iUS-based brain tumor segmentation framework with boundary-overlay visualization to support surgical assistance.

We propose a multi-scale attention-based tumor segmentation framework. Volumetric ultrasound features are extracted by a 3D convolutional backbone and hierarchically aggregated across scales. Cascaded attention refines tumor-relevant features, followed by multi-scale context aggregation to improve robustness to tumor size and appearance variations. The predicted mask is converted into a boundary overlay for intuitive iUS visualization. The method was evaluated on the public RESECT dataset and illustrated on two external clinical cases. A segmentation-assisted interpretation assessment examined whether algorithm-derived overlays improve junior neurosurgeon annotation consistency relative to a senior expert reference.

On the RESECT dataset, the proposed framework showed segmentation performance comparable to established volumetric baselines while providing a more sensitive tumor-region detection profile. Precision, specificity, and surface-distance metrics allowed the sensitivity gain to be interpreted together with false-positive behavior and surface agreement. Ablation studies supported the contribution of cascaded attention and multi-scale contextual aggregation. In the interpretation assessment, mean junior-senior agreement improved from 76.56% DSC without algorithm assistance to 90.60% DSC after reviewing the algorithm-derived overlay.

The proposed iUS-based tumor segmentation framework provides segmentation-derived boundary-overlay visualization for intraoperative interpretation. The results support its potential as a computer-assisted intraoperative interpretation tool, while further validation on larger multi-center cohorts, multiple observers, and multiple intraoperative time points remains necessary.