Our recent research in volumetric biomedical image segmentation for blood vessels has focused on improving both segmentation accuracy and computational efficiency, particularly for high-resolution microscopy data. Architectures such as 3D U-Net have become a widely adopted standard due to their ability to model hierarchical spatial features in three-dimensional volumes. Over time, more advanced variants have achieved higher performance; however, these improvements are often accompanied by significantly increased parameter counts and computational cost, making them less practical for large-scale or resource-constrained applications.
This trade-off is especially pronounced in high-resolution 3D microscopy imaging, where the input volumes are large and memory-intensive. Tasks such as vascular segmentation are inherently challenging, as these blood vessels are thin, low-contrast, and of complex structures with varying orientations. Capturing such characteristics in three dimensions is therefore important, but significantly more difficult than 2D settings, especially when it comes to resource-saving models. In addition, suitable datasets are limited in availability. High-quality volumetric microscopy data is very difficult and expensive to acquire, even for mouse brains and more so for humans. The annotations require significant expert effort, especially when one brain slice consists of over 50 million pixels, and the whole brain consists of thousands of slices. As a result, our study relies on a private dataset that consists of only a small annotated block from a single brain sample, as there are currently no available public datasets. While this allows detailed analysis of complex biological structures, it creates smaller datasets and fewer labelled samples, a problem we aim to solve with the 3D-GUSL.
To address these constraints, 3D-GUSL adopts a feed-forward U-shaped design that avoids backpropagation while preserving multi-scale spatial information. The pipeline operates hierarchically across resolution levels, where local 3D neighbourhoods are first transformed into structured feature representations using fixed kernels. These features are then compacted through PCA-based dimensionality reduction, followed by discriminative feature selection and lightweight regression using tree-based models. At each level, residual information is propagated to refine predictions, allowing the model to progressively capture both coarse and fine vessel structures. This design enables 3D-GUSL to achieve improved performance over the 3D U-Net baseline while maintaining a significantly lower parameter count and computational cost, making it well-suited for high-resolution volumetric segmentation under strict resource constraints. While running results on an unannotated test set, our collaborators found that visual analysis of unlabeled raw test images from other samples without ground truth was of high quality, but we still lack ground truth for a proper numeric evaluation.
