Predicting methylation class from diffusely infiltrating adult gliomas using multimodality MRI data.

Area of Science:

• Neuro-oncology
• Radiomics
• Machine Learning

Background:

• Radiogenomic studies traditionally rely on specific genetic alterations like IDH-mutation status and 1p19q deletion in adult diffuse gliomas.
• Current radiogenomic approaches have limitations in generalizing to tumor types lacking these recurrent genetic alterations.
• Tumors possess distinct DNA methylation patterns, enabling classification into stable methylation classes even without recurrent mutations.

Purpose of the Study:

• To establish the principle that DNA methylation class can serve as a predictive feature in radiogenomic modeling for brain tumors.
• To investigate the feasibility of using DNA methylation class for machine learning-based prediction from MRI data.

Main Methods:

• A custom DNA methylation-based classification model was employed to assign molecular classes to diffuse gliomas within The Cancer Genome Atlas (TCGA) dataset.
• Machine learning models were developed and validated to predict tumor methylation families or subclasses using multisequence MRI data.
• Predictions were made using either extracted radiomic features or directly from MRI images.

Main Results:

• Models using radiomic features achieved over 90% accuracy in predicting IDH-glioma and GBM-IDHwt methylation families and subclasses.
• Direct MRI image classification models showed average accuracies of 80.6% for predicting methylation families.
• Models differentiating IDH-mutated astrocytomas from oligodendrogliomas and glioblastoma subclasses reached accuracies of 87.2% and 89.0%, respectively.

Conclusions:

• MRI-based machine learning models effectively predict brain tumor methylation classes.
• This radiogenomic approach holds the potential to generalize to a wider range of brain tumor types.
• The findings expand the scope for developing radiomic and radiogenomic models across diverse brain tumors.