Trans-Unet: AI predicts brain folding patterns with high-fidelity 3D point-cloud learning

A new AI framework called Trans-Unet aims to solve the challenge of learning high-fidelity features from 3D point-cloud data, specifically for predicting brain folding morphology. The approach, detailed in a preprint on arXiv, first transforms 3D point-clouds into a 2D grid domain to reduce computational cost and the curse of dimensionality. It then uses a U-shaped hybrid architecture that integrates convolutional neural networks (CNNs) for local, hierarchical features and self-attention mechanisms for global, long-range dependencies. The dataset consists of 3D point-clouds with 40,401 points for brain surface patches and 2,382 points for fiber information, generated by a large-scale finite element model.

In experiments, Trans-Unet was applied to predict brain surface folding from an initial state (state 0 or states 0-2) to the final state (state 3). The results show that Trans-Unet achieves high-resolution predictions of brain patch growth, surpassing existing methods in both fidelity and accuracy. This work has implications for understanding neurodevelopmental disorders and could be extended to other biomedical imaging tasks requiring fine-grained 3D surface reconstruction. The paper is authored by Geran Zhao, Xiaotian Li, Poorya Chavoshnejad, Mir Jalil Razavi, Akbar Solhtalab, Lijun Yin, and Guifang Fu, and is available on arXiv.