Yao, Tianyuan; Li, Zhiyuan; Kim, Michael E.; Gao, Chenyu; Lionts, Marilyn; Kanakaraj, Praitayini; Archer, Derek B.; Schilling, Kurt; Beason-Held, Lori; Resnick, Susan; Hohman, Timothy J.; Jefferson, Angela L.; Landman, Bennett A.; Huo, Yuankai. (2027).听.听Medical Image Analysis, 115, 104336.听
Diffusion-weighted MRI (dMRI) provides valuable information about brain structure and neural connections, but its complex, high-dimensional data make it difficult to analyze with standard deep learning methods. We developed a Unified Polyhedral Representation (UPER), a deep learning framework designed to analyze diverse dMRI datasets within a single model. UPER organizes complex spatial and diffusion information into a simplified polyhedral representation and uses transformer-based learning to handle different imaging configurations. The model was trained using 2,800 scans from 17 datasets and performed better than traditional deep learning approaches across several analysis tasks. UPER provides a flexible, unified approach for analyzing dMRI data that could reduce the need to develop separate models for different datasets and imaging settings.
Fig. 1.听Diffusion MRI uniquely captures聽both spatial (X, Y, Z)听补苍诲听diffusion weightings (b-value, b-vector听(饾溄,饾潜)), resulting in a high-dimensional image modality. The b-vectors at each b-value can be visualized as points distributed on concentric spheres, where each sphere corresponds to one diffusion shell with approximately constant diffusion weighting. Here, the b-value and b-vector听(饾溄,饾潜)聽provide three degrees of freedom for the diffusion weighting under conventional single-diffusion-encoding (Stejskal鈥揟anner) acquisitions, as used in this work; additional degrees of freedom arise in tensor-valued encoding schemes, where the shape of the b-tensor (e.g., linear, planar, or spherical encoding) is also varied. Compared to other 3D medical imaging modalities, dMRI includes additional diffusion sensitization and directional information, providing richer tissue characterization but posing fundamental challenges for machine learning modeling. Mathematically, the dMRI signal can be represented as a function聽饾憜:饾晩2鈫掆劃饾惖, where聽饾惖聽is the number of diffusion weighting strengths (b-values), where each b-value defines a聽diffusion shell, corresponding to a set of measurements sampled on the unit sphere with approximately constant diffusion weighting.