Data-driven classification of tissue water populations by massively multidimensional diffusion-relaxation correlation MRI

Omar Narvaez¹Maxime Yon^1,2Raimo A. Salo¹Jenni Kyyriäinen¹Melina Estela¹Ekaterina Paasonen^1,3Ville Leinonen^4,5Juhana Hakumäki^4,6Frederik Bernd Laun⁷Daniel Topgaard²Alejandra Sierra^1*

• ¹A.I. Virtanen Institute for Molecular Sciences, University of Eastern Finland, Kuopio, Finland
• ²Department of Chemistry, Lunds Universitet, Lund, Sweden
• ³Neurocenter, Kuopion yliopistollinen sairaala, Kuopio, Finland
• ⁴Institute of Clinical Medicine, University of Eastern Finland, Kuopio, Finland
• ⁵Neurocenter Neurosurgery, Kuopion yliopistollinen sairaala, Kuopio, Finland
• ⁶Department of Clinical Radiology, Kuopion yliopistollinen sairaala, Kuopio, Finland
• ⁷Institute of Radiology, Universitatsklinikum Erlangen, Erlangen, Germany

Massively multidimensional diffusion-relaxation correlation MRI provides detailed information on tissue microstructure by analyzing water populations at a sub-voxel level. This method correlates frequency-dependent tensor-valued diffusion MRI with longitudinal and transverse relaxation rates, generating nonparametric D(ω)-R1-R2 distributions. Traditionally, D(ω)-R1-R2 distributions are separated using manual binning of the diffusivity and anisotropy space to differentiate white matter (WM), gray matter (GM), and free water (FW) in brain tissue. However, while effective, this approach oversimplifies complex tissue fractions and does not fully utilize all available diffusion-relaxation parameters. In this study, we implemented an unsupervised clustering approach to automatically classify WM, GM, and FW and explore additional water populations using all components in the D(ω)-R1-R2 distributions on ex vivo and in vivo rat brain, and in vivo human brain. Results showed that a basic separation of WM, GM, and FW is possible using unsupervised clustering even under different multidimensional diffusion-relaxation protocols of rat brain and human brain. Additionally, when there is high frequency-dependent diffusion range, it is possible to obtain a cluster characterized by restriction localized in specific high cell density regions such as the dentate gyrus and cerebellum of rat brain. These findings were compared with rat histological sections of myelin and Nissl stainings. We demonstrated that unsupervised clustering of diffusion-relaxation MRI data can reveal tissue complexity beyond traditional WM, GM, and FW segmentation in rat and human brain without parameter assumptions. The unsupervised cluster approach could be used in other body parts (e.g., prostate and breast cancer) without requiring pre-defined bin limits. Furthermore, the characterization of the clusters by diffusivities, anisotropy, and relaxation rates can provide a better understanding of the subtle changes in different cellular fractions in tissue-specific pathologies.