SFCNEXT: A Simple Fully Convolutional Network for Effective Brain Age Estimation with Small Sample Size

Yu Fu, Yanyan Huang, Shunjie Dong, Yalin Wang, Tianbai Yu, Meng Niu, Cheng Zhuo

Introduction

Estimating "brain age"—the predicted chronological age of an individual based on neuroimaging data—has emerged as a vital biomarker for identifying early signs of neurodegenerative diseases. A significant "brain age gap" (the difference between predicted and actual age) often correlates with conditions such as Alzheimer’s disease. However, state-of-the-art deep learning models typically require massive datasets, such as the UK Biobank with tens of thousands of MRIs, to achieve high accuracy. In many clinical settings, researchers only have access to small cohorts of a few hundred scans, where large models often overfit.

To address this, researchers have developed SFCNeXt, a lightweight, fully convolutional network designed for high-performance brain age estimation specifically in small-data scenarios.

Methodology

Single Pathway Encoded ConvNeXt (SPEC)

The core of SFCNeXt is a streamlined 3D ConvNeXt backbone. Unlike standard architectures that use many blocks per stage, SFCNeXt employs a (1, 1, 3, 1) configuration across four stages to maintain efficiency. A critical innovation is the use of overlapped down-sampling, which preserves fine-grained morphological details that are often lost during traditional non-overlapping patch division.

Overview of the SFCNeXt architecture, including the 3D ConvNeXt backbone, overlapped down-sampling, and the Conformer-based feature fusion module.

The model integrates sex features via a dedicated branch. These features are fused with the MRI-derived features using a Conformer encoder—a block that combines the local feature extraction of convolutions with the global context capabilities of Transformers—followed by a multilayer perceptron (MLP) for final age prediction.

Hybrid Ranking Loss (HRL)

Beyond architectural improvements, SFCNeXt introduces a novel Hybrid Ranking Loss (HRL). While standard models minimize Mean Squared Error (MSE), HRL incorporates a differentiable "soft ranking" component. This forces the model to learn relative age relationships (e.g., "Subject A is younger than Subject B") within each training batch. This approach effectively increases the information density per sample, making the model more robust against overfitting on small datasets.

Experimental Results

The model was evaluated on a cohort of 779 healthy subjects (aged 19–72) from 11 sites in Japan. In the context of deep learning for MRI, this is considered a small-sized dataset.

SFCNeXt demonstrated superior performance compared to established lightweight models like 3D ResNet and SFCN:

  • Accuracy: Achieved a Mean Absolute Error (MAE) of 3.683 years, outperforming the best SFCN variant (4.039 years).
  • Correlation: Reached a Pearson Correlation Coefficient (PCC) of 0.859, indicating a strong linear relationship between predicted and actual ages.
  • Efficiency: With a GPU memory footprint of approximately 6.9 GB, the model is suitable for deployment on consumer-grade hardware.
Comparison of brain age estimation performance across 3D ResNet, SFCN, and SFCNeXt. SFCNeXt shows the tightest alignment with the identity line and the lowest MAE.

Conclusion

SFCNeXt proves that deep depth is not a prerequisite for high accuracy in medical imaging if the architecture and loss functions are optimized for the data scale. By combining a lightweight ConvNeXt backbone with ranking-based supervision, SFCNeXt provides a powerful tool for brain age estimation in clinical research where data is scarce. Future work will focus on validating the model's sensitivity to specific neurodegenerative pathologies.

Summary updated

BibTeX
@inproceedings{fu2023sfcnext,
  author = {Yu Fu and Yanyan Huang and Shunjie Dong and Yalin Wang and Tianbai Yu and Meng Niu and Cheng Zhuo},
  title = {{SFCNEXT: A Simple Fully Convolutional Network for Effective Brain Age Estimation with Small Sample Size}},
  booktitle = {ISBI},
  year = {2023},
  doi = {10.1109/isbi53787.2023.10230370},
}
Figure