Machine Learning Applications in Glioma Imaging and Classification

Machine learning has emerged as a transformative approach to enhance the non-invasive characterisation of gliomas, guiding diagnosis, prognosis and treatment planning. Modern pipelines employ a range of strategies—from radiomics to deep neural networks—to mine quantitative features from multimodal MRI and PET scans. Radiomic workflows extract high-dimensional texture, histogram and shape descriptors that capture tumour heterogeneity and correlate with histological grade and genetic markers. Deep learning advances, particularly convolutional neural networks, automate feature extraction and classification, yielding robust prediction of isocitrate dehydrogenase (IDH) mutation status, 1p/19q codeletion and MGMT promoter methylation. More recently, multi-modal frameworks have integrated focal tumour imagery with geometric tumour descriptors and global brain network connectivity, facilitated by self-supervised learning and attention mechanisms, to improve genotype prediction. End-to-end multi-task deep learning models now concurrently predict tumour grade, molecular alterations and patient outcomes, while providing interpretable associations with oncogenic pathways and immune profiles. Collectively, these methods offer reproducible, high-accuracy tools that may reduce reliance on invasive biopsy, support personalised medicine and inform radiotherapy planning.

Research from Nature Portfolio

Recent studies have refined radiomics and deep learning models to predict key molecular alterations in glioma subtypes. Deep learning-based radiomics was applied to low-grade glioma, extracting features from late convolutional layers to predict IDH1 mutation with an AUC exceeding 95% across multisequence MRI data. A complementary approach incorporated lesion location information into conventional radiomics models for grade II/III tumours, combining spatial and texture features to enhance diagnostic accuracy for IDH and TERT promoter mutations, achieving validation accuracies up to 87% when lesion topology was included.

Machine Learning Applications in Glioma Imaging and Classification publication trend

The graph below shows the total number of articles in machine learning applications in glioma imaging and classification across all publications each year (not limited to Nature Index journals).

Technical terms

Convolutional neural network (CNN): A deep learning architecture employing learnable filters to capture spatial hierarchies in imaging data.

Radiomics: The extraction and analysis of large numbers of quantitative imaging features to describe tumour phenotype.

Isocitrate dehydrogenase (IDH) mutation: A genetic alteration affecting metabolic enzymes, serving as a key prognostic and diagnostic biomarker in gliomas.

1p/19q codeletion: The combined loss of chromosomal arms 1p and 19q, characteristic of oligodendrogliomas and associated with improved therapeutic response.

Deep prognosis score (DPS): A risk stratification index derived from a neural network’s outputs, correlating with survival probability and underlying biology.

References

1. Multi-Modal Learning for Predicting the Genotype of Glioma. IEEE Transactions on Medical Imaging (2023).
2. Deep Learning based Radiomics (DLR) and its usage in noninvasive IDH1 prediction for low grade glioma. Scientific Reports (2017).
3. Lesion location implemented magnetic resonance imaging radiomics for predicting IDH and TERT promoter mutations in grade II/III gliomas. Scientific Reports (2018).
4. Biologically interpretable multi-task deep learning pipeline predicts molecular alterations, grade, and prognosis in glioma patients. npj Precision Oncology (2024).