Predicting Remaining Survival of Glioblastoma Patients with Radiomics Analysis Based on ¹⁸F-DOPA PET Images

Area of Science:

• Neuro-oncology
• Radiology
• Machine Learning

Background:

• Accurate glioblastoma prognosis and monitoring are vital for timely salvage treatment.
• Differentiating tumor progression from treatment effects in conventional imaging remains a challenge.
• ¹⁸F-DOPA PET radiomics offers a potential solution for improved outcome prediction.

Purpose of the Study:

• To correlate radiomics features from serial ¹⁸F-DOPA PET scans with patient outcomes.
• To develop machine learning models for predicting remaining survival (RS) in glioblastoma patients.
• To assess the utility of these models in monitoring tumor changes and guiding treatment decisions.

Main Methods:

• ¹⁸F-DOPA PET images were analyzed from patients with wild-type IDH/unmethylated MGMT glioblastoma post-radiation therapy.
• Quantitative radiomics features were extracted from high-uptake regions.
• Machine learning algorithms and manifold learning were employed to associate imaging features with RS.

Main Results:

• Machine learning models achieved 81-83% ROC_AUC in predicting RS on an independent test set.
• A novel RS map was developed for monitoring tumor alterations via serial ¹⁸F-DOPA PET.
• The RS map demonstrated superior sensitivity and correlation with survival compared to RANO criteria.

Conclusions:

• Machine learning models using follow-up ¹⁸F-DOPA PET images can predict survival in glioblastoma patients.
• This imaging approach aids in patient stratification and salvage treatment selection.
• It also assists in differentiating treatment effects from true tumor progression, enhancing clinical management.