Abstract
Introduction: Gliomas can alter white matter pathways, causing structural deformation and functional impairments. However, there is no method to map these structural alterations across the whole brain to support diagnosis, prognosis, and treatment planning. To address this gap, we propose a novel framework based on a convolutional autoencoder that learns geometry-preserving streamline embeddings for mapping structural connectivity anomalies.
Methods: We trained and validated the autoencoder on 100 million synthetic streamlines and evaluated the anomaly-mapping pipeline using data from 118 healthy subjects. We analyzed how different tractography algorithms affect anomaly scores and demonstrated its feasibility in five glioma patients using multi-shell and single-shell diffusion MRI.
Results: Latent-space distances strongly associated with native-space geometric distances (R
2
= 0.82-0.98), supporting preserved streamline geometry in the learned embeddings. Anomaly scores varied with both the input and reference tractography methods. Patient anomaly maps showed elevated scores near gliomas and extending distally along affected pathways. These whole-brain anomaly maps were generated within minutes.
Conclusion: The framework offers a practical way to quantify and visualize geometric deviations in whole-brain tractograms, providing novel insights beyond current network and microstructural analyses, with relevance to clinical applications where white matter deformation is pertinent.