Abstract
Understanding the complex 3D network of brain blood vessels is vital for researching cognitive disorders linked to vascular dysfunction. However, manual segmentation of high-resolution 3D multiphoton microscopy (MPM) images is a significant bottleneck. While convolutional neural networks (CNNs) have automated this process, they often fail to maintain the structural connectivity of vessels, especially in noisy, deep-tissue images. We present a novel topological encoding CNN that utilizes persistent homology to ensure structural integrity. Our model outperforms existing state-of-the-art methods in Dice coefficient and produces segmentations that better mimic the topological characteristics of manual ground truth.
Introduction
Mapping the brain's dense vasculature is essential for studying neurodegenerative diseases like Alzheimer’s. 3D multiphoton microscopy provides the necessary sub-micron resolution, but the resulting data is massive and difficult to process. Standard CNNs, while efficient, often produce "broken" vessels or artificial holes in regions with low signal-to-noise ratios. These topological errors occur because traditional loss functions, like Dice loss or cross-entropy, focus on local pixel-level accuracy rather than global structural connectivity.
Methodology
The proposed approach addresses these limitations by incorporating Persistent Homology (PH)—a tool from algebraic topology—directly into the training process.
Architecture and Loss Function
The model builds upon the DeepVess architecture, utilizing 3D convolutional layers to capture spatial context. The key innovation is a hybrid loss function that balances three distinct objectives:
- Cross-Entropy: For precise voxel-wise classification.
- Generalized Dice Loss: To maximize the overall overlap between predicted and ground-truth volumes.
- Topological Loss: A differentiable term based on the Wasserstein-1 distance between persistence diagrams. This term penalizes the creation of disconnected components or artificial loops.
Computational Efficiency
Calculating PH is computationally intensive. To make training feasible, the authors implemented a sampling strategy where the topological loss is only calculated on 10% of the samples in each mini-batch. Furthermore, by optimizing the field of view to a $21 \times 21 \mu m^2$ region, the target topology could be simplified to a single connected component with no holes, streamlining the mathematical constraints.
Key Results
The model was evaluated using the DeepVess dataset, which contains manually labeled 3D MPM images from 24 mice. The results demonstrate significant improvements in both quantitative metrics and qualitative structural integrity.
Quantitative Performance
The topological encoding model achieved a Dice coefficient of 86.5%, significantly outperforming the original DeepVess (81.6%) and 3D U-Net models. It also showed greater robustness to imaging depth; while standard models' performance degrades as signal quality drops deeper in the brain, the proposed model maintained consistent accuracy.
Qualitative Improvements
Beyond numerical scores, the model produced segmentations with far fewer discontinuities. By penalizing "broken" vessels, the topological loss ensures that the resulting 3D maps are suitable for downstream applications, such as blood flow simulations, which require continuous vessel paths.
Conclusion
By bridging the gap between deep learning and algebraic topology, this research provides a robust framework for segmenting complex biological structures. The use of persistent homology as a regularizer ensures that the network "understands" the fundamental connectivity of the objects it segments. Future work will explore generalizing this approach to other tubular structures, such as neuronal networks and pulmonary airways.
Source: A topological encoding convolutional neural network for segmentation of 3D multiphoton images of brain vasculature using persistent homology (CVPR 2020 Workshop). Code available on GitHub.
Summary updated