An Ultrasound Signal-Guided Two-Stage Weakly Supervised Network for Intraoperative Glioma Localization and Infiltration Boundary Recognition Using a Nude Mouse Model.

Abstract

Objective: Surgical resection is the standard treatment for glioma. While gross tumor regions can be identified, microscopic infiltration is often elusive without histopathology. Developing real-time techniques to approximate gold-standard boundaries intraoperatively is crucial for surgical accuracy and patient outcomes.

Methods: We propose a ultrasound signal-guided two-stage spatiotemporal feature-aware weak supervision network for glioma infiltration boundaries, utilizing nude mouse pathological annotations as reference standards. In Stage 1, a spatio-temporal feature extraction module generates pseudo-boundary masks through multi-constraint learning, effectively translating the ultrasound radio frequency signal into anatomically plausible boundary probability distributions. Building upon these masks as dynamic anatomical priors, Stage 2 establishes cross-task reinforcement between tumor classification and boundary refinement in an end-to-end architecture. This cross-task synergy enhances localization accuracy with limited labels, enabling annotation-efficient and real-time intraoperative localization.

Results: Trained on 3400 intraoperative ultrasound frames (1400 tumor/2000 normal) with frame-level signal labels, the model was evaluated on a test set comprising 680 nude mouse frames (280 tumor/400 normal) using pathological annotations. For tumor/normal frame differentiation, the model achieved an accuracy of 0.985, AUC of 0.990, sensitivity of 1.000, and specificity of 0.975. Boundary recognition yielded a Dice coefficient of 0.814, intersection over union of 0.690, Hausdorff distance of 25.088, and average surface distance of 8.359 against histopathology.

Conclusion: Our method enabled accurate tumor localization with infiltration boundaries and tumor sizes closely matching the pathological gold standard, outperforming preoperative MRI. This approach offers a reliable solution for intraoperative ultrasound-assisted tumor localization, laying the foundation for clinical validation.