Optimizing Brain Age Estimation: Transfer Learning for Enhanced Model Adaptability in MRI Datasets
Poster No:
897
Submission Type:
Abstract Submission
Authors:
Tzu-Yi Yang1, Chen-Yuan Kuo2,3, Pei-Lin Lee4,5, Lee Han-Jui6,7, Kun-Hsien Chou8,3, Lin Ching-Po9,3,10
Institutions:
1National Yang Ming Chiao Tung University, Taipei, Taipei, 2Department of Neurology, Neurological Institute, Taipei, Taiwan, 3Institute of Neuroscience, National Yang Ming Chiao Tung University, Taipei, Taiwan, 4National Yang Ming Chiao Tung University, Taipei, Taiwan, 5Center for Healthy Longevity and Aging Sciences, National Yang Ming Chiao Tung University, Taipei, Taiwan, 6Department of Radiology, Taipei Veterans General Hospital, Taipei, Taiwan, 7School of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan, 8Brain research center, National Yang Ming Chiao Tung University, Taipei, Taiwan, 9Department of Education and Research, Taipei City Hospital, Taipei, Taiwan, 10Brain Research Center, National Yang Ming Chiao Tung University, Taipei, Taiwan
First Author:
Co-Author(s):
Chen-Yuan Kuo
Department of Neurology, Neurological Institute|Institute of Neuroscience, National Yang Ming Chiao Tung University
Taipei, Taiwan|Taipei, Taiwan
Department of Neurology, Neurological Institute|Institute of Neuroscience, National Yang Ming Chiao Tung University
Taipei, Taiwan|Taipei, Taiwan
Pei-Lin Lee
National Yang Ming Chiao Tung University|Center for Healthy Longevity and Aging Sciences, National Yang Ming Chiao Tung University
Taipei, Taiwan|Taipei, Taiwan
National Yang Ming Chiao Tung University|Center for Healthy Longevity and Aging Sciences, National Yang Ming Chiao Tung University
Taipei, Taiwan|Taipei, Taiwan
Lee Han-Jui
Department of Radiology, Taipei Veterans General Hospital|School of Medicine, National Yang Ming Chiao Tung University
Taipei, Taiwan|Taipei, Taiwan
Department of Radiology, Taipei Veterans General Hospital|School of Medicine, National Yang Ming Chiao Tung University
Taipei, Taiwan|Taipei, Taiwan
Kun-Hsien Chou
Brain research center, National Yang Ming Chiao Tung University|Institute of Neuroscience, National Yang Ming Chiao Tung University
Taipei, Taiwan|Taipei, Taiwan
Brain research center, National Yang Ming Chiao Tung University|Institute of Neuroscience, National Yang Ming Chiao Tung University
Taipei, Taiwan|Taipei, Taiwan
Lin Ching-Po
Department of Education and Research, Taipei City Hospital|Institute of Neuroscience, National Yang Ming Chiao Tung University|Brain Research Center, National Yang Ming Chiao Tung University
Taipei, Taiwan|Taipei, Taiwan|Taipei, Taiwan
Department of Education and Research, Taipei City Hospital|Institute of Neuroscience, National Yang Ming Chiao Tung University|Brain Research Center, National Yang Ming Chiao Tung University
Taipei, Taiwan|Taipei, Taiwan|Taipei, Taiwan
Introduction:
Brain age, derived from neuroimaging data, represents an individual's biological age and has emerged as a valuable signature for brain health [1, 2]. Recent advancements in machine learning (ML) and deep learning (DL) have significantly improved the accuracy of brain age prediction models [3]. However, brain age estimation in clinical applications remains limited due to systematic prediction biases arising from cross-site effects, including scanner differences, magnetic field strengths, and other contributing factors. These biases compromise the models' generalizability and robustness in applications across diverse datasets [3]. Therefore, this study utilized DL to develop a pre-trained brain age model and applied fine-tuning for transfer learning to adapt to new clinical data, enhancing predictive accuracy and adaptability. Additionally, we systematically investigated the impact of fine-tuning data size on model performance, offering quantitative insights for clinical applications.
Methods:
T1-weighted MRI data were collected from five sites in Taiwan, divided into a training set (N=1,400; age=18-92; 762 M/638 F) for brain age prediction model construction and a test set (N=290; age=18-92; 148 M/142 F) for brain age prediction evaluation. Additionally, we collected data from 455 healthy individuals at the Taipei Veterans General Hospital (VGH) as the clinical dataset, which was further divided into an out-of-sample set (N=305; age=20-82; 189 M/116 F) for transfer learning and a clinical test set (N=150; age=23-80; 88 M/62 F) to evaluate prediction generalizability.
All T1 images underwent standardized preprocessing, including N4 bias field correction, brain extraction and min-max normalization. 3D ResNet18 and 3D DenseNet121 [4] [5] baseline prediction performance was evaluated through 5-fold cross-validation on the training set and the final training model was used to estimate brain age in the test set and clinical set. DL models were trained with a 16-batch size, 150 epochs, the Adam optimizer and MAE (L1 Loss) as the loss function. Early stopping was applied to reduce the risk of overfitting.
Fine-tuning was employed for transfer learning, retaining the same hyperparameter settings as used during baseline model training, but reducing the number of epochs to 50. To evaluate the impact of fine-tuning sample size on model performance, experiments were conducted with varying sample sizes (10, 25, 50, 75, 100, 125, 150, 200 and 250 scans) from the out-of-sample dataset. For each sample size, the clinical test set was kept constant, and fine-tuning was repeated five times with randomly sampled out-of-sample datasets. Model performance was assessed using two metrics: mean absolute error (MAE) and the coefficient of determination (R²).
All T1 images underwent standardized preprocessing, including N4 bias field correction, brain extraction and min-max normalization. 3D ResNet18 and 3D DenseNet121 [4] [5] baseline prediction performance was evaluated through 5-fold cross-validation on the training set and the final training model was used to estimate brain age in the test set and clinical set. DL models were trained with a 16-batch size, 150 epochs, the Adam optimizer and MAE (L1 Loss) as the loss function. Early stopping was applied to reduce the risk of overfitting.
Fine-tuning was employed for transfer learning, retaining the same hyperparameter settings as used during baseline model training, but reducing the number of epochs to 50. To evaluate the impact of fine-tuning sample size on model performance, experiments were conducted with varying sample sizes (10, 25, 50, 75, 100, 125, 150, 200 and 250 scans) from the out-of-sample dataset. For each sample size, the clinical test set was kept constant, and fine-tuning was repeated five times with randomly sampled out-of-sample datasets. Model performance was assessed using two metrics: mean absolute error (MAE) and the coefficient of determination (R²).
Results:
Our results showed that DenseNet-based brain age model had better prediction performance on training set (MAE=4.359, R²=0.909) and test set (MAE=3.794, R²=0.932) compared to the ResNet-based brain age model (training: MAE=5.704, R²=0.857; test: MAE=5.551, R²=0.869). On the VGH dataset, the prediction performance of two DL-based brain age model was similar (DenseNet: MAE=7.811, R²=0.380; ResNet: MAE=7.244, R²=0.447) (Figure 1).
Additionally, both DL-based models demonstrated improved performance on the clinical test set as the amount of fine-tuning data increased, with the most efficient performance observed at a fine-tuning data size of 75 (Figure 2).
Additionally, both DL-based models demonstrated improved performance on the clinical test set as the amount of fine-tuning data increased, with the most efficient performance observed at a fine-tuning data size of 75 (Figure 2).
Conclusions:
In summary, we provided empirical evidence that DenseNet model achieved good performance. Our study demonstrated that transfer learning techniques could enhance predictive accuracy by adapting pre-trained models to new datasets. Furthermore, our analysis revealed that substantial enhancements in model accuracy can be achieved with limited data sample sizes. Ultimately, this study highlighted transfer learning as an effective and potential strategy for enhancing the clinical applicability of brain age models.
Lifespan Development:
Aging 1
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Keywords:
Aging
Modeling
MRI
STRUCTURAL MRI
Other - Deep learning, Transfer learning
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
2. Kumari, LK Soumya, and R. Sundarrajan. "A review on brain age prediction models." Brain Research 1823 (2024): 148668.
3. Dinsdale, Nicola K., Mark Jenkinson, and Ana IL Namburete. "Deep learning-based unlearning of dataset bias for MRI harmonisation and confound removal." NeuroImage 228 (2021): 117689.
4. He, Kaiming, et al. "Deep residual learning for image recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
5. Huang, Gao, et al. "Densely connected convolutional networks." Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.