Summary: A new machine-learning algorithm accurately distinguishes lower-grade gliomas from glioblastoma using clinically relevant MRI measurements of tumor location and volume.
Source: Yale
An estimated 18,000 people in the United States died from brain and spinal cord tumors in 2020. To help clinicians more quickly and reliably distinguish tumor severity, an international research team led by Dr. Murat Günel, Chair of Neurosurgery at Yale School of Medicine, and Nixdorff-German Professor of Neurosurgery, developed a machine learning model that analyzes MRI features of brain tumors. The model learns patterns in tumor appearance, location, and component volumes to predict whether a tumor is a lower-grade glioma or a glioblastoma, supporting more accurate staging and diagnosis.
To evaluate their approach, the investigators used MRI data from 229 patients with tumors spanning the clinical spectrum from lower-grade gliomas (slow-growing tumors arising from glial cells) to glioblastomas (highly aggressive grade IV tumors). The data were drawn from a public imaging repository and curated by board-certified neuroradiologists to ensure clinically relevant case selection for model training and validation.
“Our machine learning models used to differentiate the tumor types were very accurate,” said Hang Cao, a medical student at Xiangya Hospital working with Dr. Günel and the lead author of the study published in European Radiology.
The study identified clear differences in tumor appearance, spatial distribution across brain regions, and relative volumes of enhancing tumor, non-enhancing tumor, and peritumoral edema. When combined into a quantitative model, these features enabled accurate discrimination between lower-grade gliomas and glioblastoma on validation datasets.
Researchers extracted tumor location and volumetric component features from the imaging collection and applied multiple statistical and machine learning techniques. They evaluated model performance using two sampling strategies: institution-based sampling and repeated random sampling (10 iterations with 70% training and 30% validation splits). Methods tested included LASSO (least absolute shrinkage and selection operator) regression and nine different machine learning algorithms, with ensemble and support vector machine approaches demonstrating the strongest results.
Key findings included high classifier accuracy and area under the receiver operating characteristic curve (AUC). For several models, validation set AUCs exceeded 0.90 and classifier accuracy surpassed 0.79, while repeat random sampling produced average validation AUCs above 0.93 and average accuracies above 0.85. The LASSO regression model, relying on features such as the percentage of tumor in the frontal lobe and volumes of enhancing and non-enhancing tumor, also reached high performance (institution-based validation AUC 0.909, accuracy 0.830), and the authors derived a formula for that best-performing LASSO model.
Although the machine learning models show robust performance in retrospective testing, the authors emphasize that clinical deployment requires standardized workflows and broader validation. Implementing such models in routine radiology would involve integration with existing software and hardware, and agreement on performance and safety standards by the scientific community and manufacturers before use in patient care.
“This work is fundamentally important to our understanding of brain tumors and a great example of the collaborative, multidisciplinary effort we use to advance the field and provide the best care to brain tumor patients,” said co-author Dr. Jennifer Moliterno, Assistant Professor of Neurosurgery at Yale School of Medicine and Clinical Program Leader of the Brain Tumor Program.
Source:
Yale
Media Contacts:
Jennifer Chen – Yale
Image Source:
The image is credited to Yale.
Original Research: Closed access
“A quantitative model based on clinically relevant MRI features differentiates lower grade gliomas and glioblastoma.” Jennifer Moliterno et al., European Radiology. DOI: 10.1007/s00330-019-06632-8.
Abstract
A quantitative model based on clinically relevant MRI features differentiates lower grade gliomas and glioblastoma
Objectives
To establish a quantitative MRI model that uses clinically meaningful tumor location and component volume features to differentiate lower-grade glioma (LGG, grades II and III) from glioblastoma (GBM, grade IV).
Methods
Tumor location and component volumes (enhancing tumor, non-enhancing tumor, peritumoral edema) were extracted from 229 cases drawn from The Cancer Genome Atlas (TCGA-LGG and TCGA-GBM). Using institution-based sampling and repeated random sampling (10 iterations, 70% training vs 30% validation), the team applied LASSO regression and nine machine learning algorithms to build and evaluate predictive models.
Results
Principal component analysis indicated that extracted MRI features could separate LGG and GBM cases. Ensemble stack modeling and support vector machine classifiers achieved the highest performance (institution-based validation set AUC > 0.900, accuracy > 0.790; repeated random sampling average validation AUC > 0.930, average accuracy > 0.850). The LASSO model using frontal lobe tumor percentage and volumes of enhancing and non-enhancing tumor achieved strong performance (institution-based validation AUC 0.909, accuracy 0.830), and the authors formulated the regression expression for this model.
Conclusions
Clinically meaningful, computer-derived MRI features of tumor location and component volumes produced models with high accuracy (validation AUC > 0.900, classifier accuracy > 0.790) for distinguishing lower-grade glioma from glioblastoma. These findings support further validation and the development of clinical standards for integrating machine learning assessments into radiologic workflows.