Rules-based Volumetric Segmentation of Multiparametric MRI for Response Assessment in Recurrent High-Grade Glioma

Area of Science:

• Radiology and Medical Imaging
• Oncology
• Artificial Intelligence in Medicine

Background:

• Accurate segmentation of brain tumor subregions in multiparametric MRI (mpMRI) is crucial for treatment response assessment.
• Current segmentation methods may lack consistency and interpretability, particularly in complex cases like high-grade glioma (HGG) and recurrent HGG (rHGG).
• Quantitative analysis of tumor components is essential for evaluating treatment efficacy in clinical trials.

Approach:

• Developed domain knowledge-based rules to assign voxels to distinct tissue types (e.g., contrast-enhancing tumor, edema) using mpMRI data.
• Validated segmentation rules on external datasets and assessed their correlation with ground truth tumor and edema volumes.
• Integrated rule-based tissue volumes into a predictive model for disease progression in a clinical trial of recurrent high-grade glioma (rHGG).

Key Points:

• Rule-based segmentation of contrast-enhancing tumor (CE) and peritumoral edema (High FLAIR) showed strong correlation with ground truth volumes (R=0.85 and R=0.87, respectively).
• A model combining treatment time and rule-based volumes achieved 82.5% accuracy in predicting progression within 30 days for rHGG patients.
• The explainable decision tree approach provides validated and consistent intratumoral tissue volumes.

Conclusions:

• Domain knowledge-based rules provide a validated and consistent method for segmenting brain tumor tissues on mpMRI.
• This approach enables reliable quantitative assessment of tumor components, crucial for monitoring treatment response.
• The developed method shows promise for predicting clinical progression in rHGG patients within clinical trial settings.