As the global aging problem intensifies, the incidence of neurodegenerative diseases (such as Alzheimer’s Disease, AD) is increasing year by year. Brain aging (Brain Aging, BA) is one of the significant risk factors for neurodegenerative diseases, but it does not completely align with chronological age (Chronological Age, CA). Traditional methods for assessing brain aging primarily rely on DNA methylation clocks. However, this method cannot directly reflect brain tissue aging because the blood-brain barrier separates blood cells from brain cells. Therefore, how to accurately assess the pace of brain aging (Pace of Brain Aging, P) through non-invasive means has become an important research topic.

This study aims to develop a model capable of quantifying the pace of brain aging using deep learning techniques and longitudinal magnetic resonance imaging (Longitudinal MRI) data, and explore its relationship with neurocognitive changes. This research not only helps in earlier identification of populations at risk of neurodegenerative diseases but also provides a scientific basis for personalized intervention strategies.

Source of the Paper

This paper was jointly completed by Chenzhong Yin, Phoebe Imms, Nahian F. Chowdhury, Nikhil N. Chaudhari, Heng Ping, Haoqing Wang, Paul Bogdan, Andrei Irimia, and others. The research team comes from various departments of the University of Southern California (USC), including the Department of Electrical and Computer Engineering, the Gerontology Research Center, and the Department of Biomedical Engineering. The paper was published in the Proceedings of the National Academy of Sciences (PNAS) on February 24, 2025.

Research Process and Details

1. Research Design

The research team developed a longitudinal model (LM) based on a three-dimensional convolutional neural network (3D Convolutional Neural Network, 3D-CNN) to estimate the pace of brain aging from longitudinal MRI data. The research is divided into several main steps:

a) Data Collection and Preprocessing

The study used MRI data from multiple databases, including the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and the UK Biobank (UKBB). A total of 3,359 cognitively normal (CN) adults aged between 47 and 88 were included in the study. Additionally, the study included an independent test set of 104 cognitively normal individuals and 140 Alzheimer’s disease patients.

MRI data underwent preprocessing, including skull stripping, motion correction, signal intensity normalization, and brain segmentation and reconstruction using Freesurfer.

b) Model Development and Training

The research team designed a 3D-CNN model, which takes the difference in MRI volumes (δi = i(t2) - i(t1)) of the same participant at two time points (baseline t1 and follow-up t2) as input. The model estimates the change in brain age (δBA = BA(t2) - BA(t1)) via regression analysis and calculates the pace of brain aging (P = δBA / δCA, where δCA is the time interval).

The model was trained using data from 2,055 cognitively normal adults, with a validation set including 1,304 cognitively normal adults. The model uses Mean Squared Error (MSE) as the loss function and is optimized using the Adam optimizer.

c) Model Testing and Comparison

The study evaluated the performance of the model on an independent test set and compared it with three existing models:
1. A 3D-CNN model based on single-time-point MRI data;
2. A Simple Fully Convolutional Network (SFCN);
3. An optimized SFCN model (SFCN-reg).

The results showed that the mean absolute error (MAE) of the longitudinal model was 0.16 years, significantly better than other models (with MAEs of 1.85 years, 2.2 years, and 2.73 years respectively).

2. Main Results

a) Estimation of the Pace of Brain Aging

The longitudinal model performed well in cognitively normal adults, accurately estimating the pace of brain aging. In Alzheimer’s disease patients, the model’s MAE was 0.50 years, still outperforming other models.

b) Association with Neurocognitive Changes

The study found that the pace of brain aging (P) was significantly associated with changes in cognitive function. For example, in the ADNI dataset, P was positively correlated with changes in the Alzheimer’s Disease Assessment Scale (ADAS) scores, indicating that the faster the brain ages, the more pronounced the decline in cognitive function.

c) Mapping of Anatomical Features

Through saliency mapping, the research team identified brain regions associated with the pace of brain aging across different genders, ages, and cognitive states. For instance, the right precentral gyrus and postcentral gyrus play a crucial role in P estimation in females, while the left transverse frontopolar gyrus and right supramarginal gyrus are more critical in males.

Conclusions and Significance

This study developed a longitudinal model based on deep learning that can accurately estimate the pace of brain aging from longitudinal MRI data and reveal its relationship with neurocognitive changes. This model not only provides a new tool for early risk assessment of neurodegenerative diseases but also lays the foundation for the development of personalized intervention strategies.

Highlights of the Study

1. Innovative Method: For the first time, 3D-CNN was used to directly estimate the pace of brain aging from longitudinal MRI data, avoiding the limitations of traditional methods that require multiple independent estimations of brain age.
   
2. High Precision and Generality: The model performed excellently in both cognitively normal adults and Alzheimer’s disease patients and can be generalized to independent test sets.
   
3. Anatomical Interpretability: Through saliency mapping, the study revealed brain regions associated with the pace of brain aging in different genders and cognitive states, providing new insights into understanding the biological mechanisms of brain aging.
   

Other Valuable Information

This study also provides important implications for future research directions, such as expanding the diversity of training samples to improve the generality of the model, and further verifying the accuracy of the model by combining other biomarkers (such as DNA methylation).