Machine learning in neuroimaging for predicting H3K27M mutations in diffuse midline gliomas: a systematic review and...

Area of Science:

• Neuro-oncology
• Medical imaging
• Artificial intelligence in medicine

Background:

• Diffuse midline gliomas (DMG) are aggressive brain tumors often characterized by H3K27M mutations.
• Accurate non-invasive prediction of H3K27M mutations is crucial for treatment planning and patient stratification.
• Current diagnostic methods can be invasive and time-consuming.

Purpose of the Study:

• To evaluate the diagnostic performance of neuroimaging-based machine learning (ML) models for non-invasive H3K27M mutation prediction in DMG.
• To compare the efficacy of MRI-based versus PET/CT-based ML models.
• To identify factors influencing model performance, such as deep learning architectures and reference standards.

Main Methods:

• A systematic review and meta-analysis was conducted following PRISMA-DTA guidelines, searching four databases up to May 2025.
• Sixteen studies involving 2,357 patients (internal validation) and 1,792 patients (external validation) were included.
• Study quality was assessed using PROBAST+AI and GRADE; bivariate random-effects models were used for meta-analysis of sensitivity, specificity, and AUC.

Main Results:

• MRI-based ML models demonstrated strong diagnostic performance with pooled sensitivity of 0.86, specificity of 0.82, and AUC of 0.91 in internal validation.
• PET/CT-based ML models showed lower performance (sensitivity 0.58, specificity 0.65, AUC 0.61).
• MRI-based ML significantly outperformed PET/CT-based ML in sensitivity and AUC (P < 0.01). Deep learning models showed superior performance over conventional ML (P = 0.01).

Conclusions:

• MRI-based ML models offer high accuracy and generalizability for non-invasive H3K27M prediction in DMG.
• MRI-based ML appears superior to PET/CT-based ML for this application.
• The use of deep learning architectures and DNA sequencing as a reference standard may further enhance predictive performance, supporting clinical utility for molecular stratification.