Integrated HSI-MRI Approach for Tumor Identification in Neurosurgery

Overview

This study presents a novel multimodal imaging system integrating preoperative MRI and intraoperative hyperspectral imaging (HSI) to enhance tumor delineation during neurosurgical procedures. The approach combines MRI's structural detail with HSI's biochemical tissue characterization, supported by machine learning classification and robust multimodal registration techniques.

Background

Accurate tumor margin identification and healthy tissue preservation are critical in neurosurgery due to the brain's complex anatomy and intraoperative changes. MRI is the standard for preoperative planning but loses accuracy intraoperatively because of brain shift. Hyperspectral imaging offers real-time biochemical insights but lacks structural context and penetration depth. Integrating these modalities addresses their individual limitations, providing comprehensive guidance during surgery.

Data Highlights

The study quantitatively evaluated three multimodal registration strategies—surface-based, landmark-based, and ArUco marker-based—using brain-like phantoms, demonstrating their accuracy and feasibility for intraoperative use. Machine learning with XGBoost was employed to classify high-dimensional hyperspectral data, effectively distinguishing tumor tissue from healthy brain tissue. MRI segmentation utilized the Segment Anything Model (SAM) guided by tumor bounding boxes to delineate tumoral regions semi-automatically.

Key Findings

• A novel visualization system was developed that integrates preoperative MRI and intraoperative HSI into a unified platform for tumor resection guidance.
• Three multimodal registration methods were implemented and quantitatively validated, confirming their accuracy and intraoperative applicability.
• Machine learning with XGBoost effectively classified hyperspectral data, capturing subtle spectral differences between tissue types.
• MRI-based semi-automatic segmentation using SAM provided reliable tumor delineation to complement HSI classification.
• The combined MRI-HSI approach offers enhanced tissue characterization beyond what either modality can provide alone.

Clinical Implications

The integrated HSI-MRI system enables more precise intraoperative tumor margin identification by combining structural and biochemical information, potentially improving surgical outcomes. Robust registration methods ensure accurate alignment of multimodal data in real time, facilitating effective image-guided neurosurgery. Machine learning classification of HSI data supports objective tissue differentiation, aiding surgeons in preserving healthy brain tissue.

Conclusion

Integrating MRI and hyperspectral imaging with advanced registration and machine learning techniques provides a promising multimodal platform for enhanced tumor identification during neurosurgery. This approach addresses limitations of individual modalities and supports improved intraoperative decision-making.

References

1. XGBoost [8] -- A scalable tree boosting system
2. Segment Anything Model (SAM) [16] -- Foundation model for image segmentation
3. Zeineldin et al. [41] -- Multimodal neuronavigation system integrating MRI and intraoperative ultrasound