SINGAPORE – Machine learning with MRI radiomics capabilities can be used to accurately assess brain aging, according to research presented at the International Society of Magnetic Resonance in Medicine (ISMRM) conference on May 6.
Presenter Eros Montin, MD New York University Grossman School of Medicine in New York City Machine learning model reported Using radiomics features from T1- and T2-weighted MR images The estimated age of adult subjects had a mean absolute error value of 4.7 years.
The findings could lead to improved clinicians' understanding of both healthy aging and brain changes due to neurodegeneration.Montin points out.
“Machine learning models that can accurately estimate brain age could have significant clinical impact,” he said.
Previous studies using structural imaging to predict brain age have shown mean absolute error values of 5 to 7 years, whereas studies combining structural and functional imaging information have shown that the mean absolute error values are between 5 and 7 years. It is shown that the error value is less than 4 years. However, functional imaging (e.g., functional MRI or diffusion-weighted MRI) may not be widely available, and predictive accuracy is only achieved when large amounts of data (i.e., 23,000 or more) are required.
This is where MRI radiomics comes into play. Sorting out quantitative features from MRI exams “is emerging as a powerful tool to improve patient outcomes and advance precision medicine,” Montin says.
“Radiomics extracts image features from specific regions of interest and uses machine learning models to correlate the features with clinical outcomes,” he explained. “Because of this, radiomics requires fewer instances than other models trained on complete images, reducing the degrees of freedom the model needs to learn.”
Montin and colleagues used aging data from the Human Connectome Project in a study involving T1.T2-weighted brain images from 716 healthy adults. The data collected 18,324 radiomics features from brain regions thought to be significantly affected during the normal aging process (bilateral hippocampus, putamen, and caudate nucleus) and used a stacking regressor machine learning model ( machine learning techniques). Montand says this is a combination of predictions from multiple estimators). The model included his eight regressors: Lasso, Random Forest, k-Nearest Neighbors, Gradient Boosting, AdaBoost, HistGradientBoostingRegressor, and MLPRegressor.
A predictive model using a stacking regressor was trained with only 20 radioactive features. This “supports the hypothesis that these three major subcortical brain regions are sufficient to provide important information for machine learning-based aging predictions,” Montin's team says. . Estimated brain age with mean absolute error value of 4.7 years.
“Our study shows that radiological features can be used to predict brain age in healthy adults, with performance comparable to that reported for models trained on significantly larger datasets available in the literature. “We're showing that,” Montin said.
Montand said the researchers plan to continue the study in patients with myosclerosis and examine whether the model and MRI radiomics show a correlation between brain age and the number of brain lesions.
