Brain age fingerprinting from MR image using multi-level information fusion networks and its application in cognitive impairment patient screening
Author Block: N. Zhao, Y. Pan, Z. Xue, F. Gao, F. Shi, D. Shen; Shanghai/CN
Purpose: The aim of this study was to develop a model for estimating brain age from MR image on a large-scale normal aging population covering entire lifespan, and to assess its potential for early screening of cognitive impairment patients.
Methods or Background: We proposed a novel approach to build a brain age prediction model in lifespan datasets using T1-weighted MR images. This approach consists of extracting three-level hierarchical information through neural networks and fusing them with the cross-attention mechanism, to capture inherent brain age fingerprinting in MR images. Specifically, the hierarchical information included: (1) brain volumes and ratios of 106 parcellations derived from a pre-trained segmentation model, (2) 2D image slices selected from specific brain regions, which potentially contain brain lesion information such as white matter hypointensities, lacunes, and perivascular spaces, (3) 3D CNN features with input of MR image.
Results or Findings: This study included 3,711 subjects aged 6-96 years from in-house datasets, with 3,372 cognitively normal (CN), 207 late MCI (LMCI), and 132 AD. Based on the proposed model, CN subjects achieved a mean absolute error of 2.72 years. Furthermore, when applying this model to cognitively impaired subjects, AD group had higher brain age gap (BAG) compared to both LMCI and CN groups (4.43 vs. 2.47 vs. -0.5 years; P < .001). Finally, combing BAG with learned age-related features as inputs of multi-layer perceptron for differentiating between CN, LMCI, and AD yielded predictive accuracy of 91% for CN vs. LMCI, 91% for CN vs. AD, and 96% for LMCI vs. AD.
Conclusion: The BAG from prediction model appears to be highly correlated with cognitive impairment and could be used for screening of cognitive impairment patients.
Limitations: The model utilises 3D high-resolution images while the extension to clinically low-resolution MRI scans should be studied.
Funding for this study: This work was supported in part by National Natural Science Foundation of China (62131015), Science and Technology Commission of Shanghai Municipality (STCSM) (21010502600), Key R&D Program of Guangdong Province, China (2021B0101420006), STI2030-Major Projects (2022ZD021 3100), The China Postdoctoral Science Foundation (Nos. BX2021333, 2021M703340), and National Key Research and Development Program of China (2022YFE02 05700). Data collection and sharing for this project was funded by Shanghai Zhangjiang National Innovation Demonstration Zone Special Funds for Major Projects (ZJ2018-ZD-012), Shanghai Pilot Program for Basic Research (JCYJ- SHFY-2022-014), and Shanghai Pujiang Program (21PJ1421400).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by Autism Brain Imaging Data Exchange (ABIDE), Attention Deficit Hyperactivity Disorder (ADHD-200), Alzheimer’s Disease Neuroimaging Initiative, Open Access Series of Imaging Studies (OASIS), Consortium for reliability and reproducibility (CoRR), Consortium of Chinese Brain Molecular and Functional Mapping, HUASHAN Hospital, and RENJI Hospital, China, approved this study.