BrainAge CNN

This project predicts chronological age from brain MRI using a convolutional neural network (CNN).

The goal is to build a simple, reproducible baseline for predicting chronological age from brain MRI, while explicitly studying how the number of input slices affects regression performance.

The aim is create an interpretable end-to-end machine learning pipeline for medical imaging, prioritizing transparency and reproducibility over state-of-the-art performance.

Problem formulation

• Task: supervised regression
• Input: structural T1-weighted brain MRI
• Output: chronological age (continuous, in years)

---

Pipeline overview

For each subject:

• Load a 3D MRI volume
• Extract a fixed number of axial slices from the central brain region
• Stack slices as channels
• Train a CNN to regress age

The project evaluates multiple configurations with different numbers of slices.

---

Dataset

• Dataset: OASIS (Open Access Series of Imaging Studies)
• Number of subjects: 351
• Target variable: chronological age

---

Slice sampling

• Axial slices are sampled uniformly between 25% and 75% of the Z-axis
• The number of slices is a configurable parameter

Experiments are run with:

• 1 slice
• 3 slices
• 5 slices
• 9 slices
• 15 slices

Input tensor shape:

• (n_slices, H, W)

This allows studying the trade-off between spatial information and model complexity.

---

Preprocessing

• MRI volumes loaded using nibabel
• Per-subject z-score normalization
• No skull stripping or spatial normalization
• Same preprocessing for all slice configurations

---

Model

• 2D convolutional neural network (PyTorch)
• Three convolutional blocks:
  • Conv2D + BatchNorm + ReLU
• Global average pooling
• Linear regression head
• Same architecture used for all experiments (only input channels change)

---

Training setup

• Supervised regression
• Loss function: Huber loss
• Optimizer: AdamW
• Train / validation / test split
• Fixed random seeds for reproducibility
• Best model selected based on validation performance

---

Evaluation

Model performance is evaluated on a held-out test set using:

• Mean Absolute Error (MAE)
• Root Mean Squared Error (RMSE)
• Bias (mean signed error)
• Pearson correlation coefficient
• R2 score
• Error distribution histograms
• Predicted vs true age scatter plots
• MAE stratified by age bins

Results are compared across different numbers of input slices.

---

Repository structure

BrainAge-CNN/ ├── data_loader.py ├── dataset.py ├── model.py ├── train.py ├── eval.py ├── plot.py ├── data_visualization.py └── README.md

---

Dataset reference

This project uses data from the OASIS (Open Access Series of Imaging Studies) dataset.

Marcus, D. S., Wang, T. H., Parker, J., Csernansky, J. G., Morris, J. C., & Buckner, R. L.
Open Access Series of Imaging Studies (OASIS): Cross-sectional MRI data in young, middle-aged, nondemented, and demented older adults.
Journal of Cognitive Neuroscience, 19(9), 1498–1507, 2007.

https://www.oasis-brains.org