Machine and Deep Learning in Hyperspectral Fluorescence-Guided Brain Tumor Surgery

Area of Science:

• Neuro-oncology
• Medical Imaging
• Machine Learning

Background:

• Malignant glioma resection is a primary neuro-oncology treatment.
• Distinguishing tumor edges, especially in infiltration zones, is challenging during surgery, even with fluorescence guidance (e.g., 5-aminolevulinic acid).
• Difficulties arise with lower-grade gliomas, tumors extending beyond MRI margins, and some high-grade tumors lacking visible fluorescence.

Purpose of the Study:

• To explore hyperspectral imaging (HSI) methods for intraoperative brain tumor margin delineation.
• To investigate the application of machine learning (ML) to HSI data for improved tissue classification and molecular status determination.
• To develop a pipeline combining classical and deep learning methods for processing HSI fluorescence data in brain tumor surgery.

Main Methods:

• Utilized ex vivo hyperspectral fluorescence imaging of brain tumor biopsies.
• Developed a pipeline for spectral data preprocessing, fluorophore abundance determination, and ML-based classification.
• Employed classical and deep learning techniques for spectral analysis and classification of tumor characteristics.

Main Results:

• Achieved high average test accuracies: 87% for tumor type, 96.1% for WHO grade, 86% for margin tissue type, and 93% for IDH mutation status.
• Demonstrated superior performance compared to prior methods, both with and without fluorescence guidance.
• Identified fluorophore abundances as direct indicators of cancerous tissue presence.

Conclusions:

• Data-driven hyperspectral imaging shows significant promise for intraoperative classification of brain tumors during fluorescence-guided surgery.
• The developed ML pipeline effectively processes HSI data to provide crucial diagnostic information.
• This approach offers a potential advancement for surgical decision-making and patient outcomes in neuro-oncology.