Imaging-Based Algorithm for the Local Grading of Glioma

Area of Science:

• Neuro-oncology
• Medical imaging
• Machine learning in medicine

Background:

• Gliomas are heterogeneous brain tumors requiring accurate grading for treatment.
• Current imaging techniques lack validation against histopathology for glioma grading.
• Predicting glioma grade from imaging can enhance diagnostic value.

Purpose of the Study:

• To estimate local glioma grade using machine learning models.
• To train models on preoperative imaging data and spatially specific tumor samples.
• To validate the predictive accuracy of imaging-based glioma grading.

Main Methods:

• Prospective clinical trial enrolling glioma patients (2013-2016).
• MR imaging with anatomic, diffusion, permeability, and perfusion sequences.
• Image-guided stereotactic biopsy and development of machine learning models (random forest) to predict World Health Organization grade.

Main Results:

• A random forest model achieved 96% accuracy (κ = 0.93) in predicting glioma grade using four specific imaging inputs.
• Conventional imaging alone resulted in 89% accuracy (κ = 0.79), with 43% of high-grade samples misclassified.
• Advanced imaging data significantly improved prediction accuracy over conventional methods.

Conclusions:

• Local glioma pathologic grade can be predicted with high accuracy using clinical imaging data.
• Advanced imaging techniques add significant value to conventional imaging for glioma grading.
• Confirmatory imaging trials are warranted to further validate these findings.