Vellmer Sebastian, Aydogan Dogu Baran, Roine Timo, Cacciola Alberto, Picht Thomas, Fekonja Lucius S
Department of Neurosurgery, Charité Universitätsmedizin Berlin, Berlin, Germany.
Cluster of Excellence, Matters of Activity, Image Space Material, Berlin, Germany.
Commun Biol. 2025 Mar 28;8(1):512. doi: 10.1038/s42003-025-07936-w.
Machine learning may enhance clinical data analysis but requires large amounts of training data, which are scarce for rare pathologies. While generative neural network models can create realistic synthetic data such as 3D MRI volumes and, thus, augment training datasets, the generation of complex data remains challenging. Fibre orientation distributions (FODs) represent one such complex data type, modelling diffusion as spherical harmonics with stored weights as multiple three-dimensional volumes. We successfully trained an α-WGAN combining a generative adversarial network and a variational autoencoder to generate synthetic FODs, using the Human Connectome Project (HCP) data. Our resulting synthetic FODs produce anatomically accurate fibre bundles and connectomes, with properties matching those from our validation dataset. Our approach extends beyond FODs and could be adapted for generating various types of complex medical imaging data, particularly valuable for augmenting limited clinical datasets.
机器学习可以增强临床数据分析，但需要大量的训练数据，而对于罕见病症来说这些数据很稀缺。虽然生成神经网络模型可以创建逼真的合成数据，如3D磁共振成像（MRI）容积数据，从而扩充训练数据集，但生成复杂数据仍然具有挑战性。纤维取向分布（FODs）就是这样一种复杂的数据类型，它将扩散建模为球谐函数，并将权重存储为多个三维容积数据。我们利用人类连接组计划（HCP）数据，成功训练了一个结合生成对抗网络和变分自编码器的α- Wasserstein生成对抗网络（α-WGAN）来生成合成FODs。我们生成的合成FODs产生了解剖结构准确的纤维束和连接组，其属性与我们验证数据集中的属性相匹配。我们的方法不仅适用于FODs，还可以适用于生成各种类型的复杂医学成像数据，这对于扩充有限的临床数据集特别有价值。
