Predicted brain age has emerged in recent years as a biomarker of brain aging [1-2], useful in several pathological conditions [1-3], when investigating brain maturation [4], and analysing lifestyles impact on aging [4].

Deep learning has become the gold standard for brain age prediction from T1-w MRI thanks to its high accuracy, compared to traditional machine learning methods.

Recently, Convolutional KANs (Kolmogorov-Arnold Networks), based on the Kolmogorov-Arnold representation theorem [5], have gained interest for their superior accuracy and efficiency compared to traditional Convolutional Neural Networks (CNN)  in tasks like classification [6], segmentation [7-9], and image generation [8].

In this study, T1-weighted 3D MRI scans from three public datasets (2129 subjects) were used to evaluate model performance, both with and without data augmentation. Data was split into 80% for training and 20% for testing. Models were trained using 5-fold cross-validation on the training set, and performance was assessed based on the median results of the ensembled folds.

CNNs achieved a correlation coefficient of R²=0.8247 and a mean absolute error (MAE) of 5.9822, whereas KAN R²=0.8787 and  MAE=5.0752. With data augmentation, CNN performance improved to R²=0.8970 and MAE=4.5885, while KAN R²=0.8985 and MAE=4.4979. Our findings demonstrate that KAN outperformed traditional CNNs in brain age prediction, however this gain is more limited in presence of data augmentation. These results highlight the potential of KAN in improving brain age prediction from T1-w MRI.

Acknowledgements:

This project was supported by the Ministry of University and Research within the Complementary National Plan PNC-I.1 “Research initiatives for innovative technologies and pathways in the health and welfare sector, D.D. 931 of 06/06/2022, PNC0000002 DARE – Digital Lifelong Prevention CUP: B53C22006440001.

Figure 1:

Scatter diagrams of estimated brain ages in the test data by different models (CNN and KAN), with and without data augmentation. In all diagrams, the red dashed line indicates the ideal estimation, and the green line is the linear regression model fitted by the estimated brain age and chronological age.

References:

[1] Soumya Kumari LK et al., Brain Res. 2024;1823:148668.

[2] Mishra S et al., IEEE Rev Biomed Eng. 2023;16:371-385.

[3] Baecker L et al., eBioMedicine. 2021;72:103600

[4] Tanveer M et al., Inf Fusion. 2023; 96:130-143

[5] Liu Z et al., arXiv preprint arXiv:2404.19756. 2024

[6] Bodner AD et al., arXiv preprint arXiv:2406.13155. 2024

[7] Tang T et al., arXiv preprint arXiv:2408.00273. 2024.

[8] Li C et al., arXiv preprint arXiv:2406.02918. 2024.

[9] Agrawal A et al., arXiv preprint arXiv:2411.11926. 2024.