Rapid estimation of structural brain disconnection from binary lesion masks using DeepDisco

Mapping how focal brain lesions disrupt structural connectivity is key to understanding cognition, behavior, and neurological outcomes. However, traditional diffusion tractography pipelines are computationally intensive, require specialized data and expertise, and remain inaccessible for many research and clinical applications. Here we present DeepDisco, an extension of the previously established Deep-Disconnectome framework by introducing an open-source, cross-platform deep-learning tool for rapid estimation of brain disconnection maps directly from binary lesion or region masks.

Built on a 3D U-Net architecture and trained on large-scale simulated connectome data, DeepDisco predicts voxelwise disconnection across major fiber systems in less than one second per lesion. It is designed to output the entire brain’s disconnected fibers, including specific disconnections within three fiber categories: association, commissural, and projection fibers. Quantitative validation against conventionally derived disconnectomes showed strong whole-brain reconstruction performance.

Fiber-class-specific reconstructions showed lower exact voxelwise correlations but preserved broader spatial organization, with structural similarity indices ranging from 0.780 to 0.829 across association, commissural, and projection fiber maps. In downstream behavioral prediction, DeepDisco-derived embeddings showed variable predictive out-of-sample performance across domains and hemispheres, reaching out-of-sample R² values up to 0.641 for motor score prediction. Notably, including fiber-class-specific disconnectomes provided domain-dependent benefits, enhancing motor outcome predictions, whereas whole-brain representations remained effective for cognitive functions.

Despite its speed, the model maintained a minimal computational footprint (