Preserving Topological Connectivity in Thin Structure Segmentation

1,218 concepts720 papers25 proposals

When deep learning models segment objects in images, they typically try to maximize pixel-level accuracy. For large, bulky objects like liver organs or buildings, missing a few pixels along the border is negligible. However, when segmenting thin, branching, or network-like structures—such as blood vessels, airways, roads, or neurons—a single-pixel mistake can be catastrophic.

If a model fails to predict just one pixel in the middle of a long capillary or a highway, the volumetric overlap remains nearly perfect, but the underlying network is severed. This phenomenon is known as topological fragmentation.

The Illusion of Volumetric Overlap

Standard semantic segmentation models are trained using pixel-independent loss functions like Cross-Entropy and evaluated using overlap metrics like the Dice coefficient or Intersection over Union (IoU). These metrics treat every pixel with equal weight, completely ignoring how pixels connect to form global structures.

As highlighted by Hao Li et al. and Stucki et al., a model can achieve a near-perfect Dice score while producing a segmentation map that is clinically or practically useless. For instance, a broken vessel prevents blood-flow simulations, and a fragmented road network breaks pathfinding algorithms (Mattyus et al.).

Traditional architectures exacerbate this problem because their downsampling operations (such as max-pooling) naturally discard fine, low-contrast details, causing thin structures to appear intermittent or broken (Song Guo et al.). Furthermore, at complex junctions where branches cross, models often lose track of continuity, leading to artificial splits or incorrect mergers (Xiaoyu Yang et al., Jin et al.).

Differentiable Topology via Persistent Homology

To force neural networks to respect connectivity, researchers have integrated concepts from algebraic topology into loss functions. The primary tool used is persistent homology, which mathematically tracks the "birth" and "death" of topological features—such as connected components ($H_0$) and loops or holes ($H_1$)—as an image is thresholded at varying levels of intensity.

The breakthrough lies in making this topological tracking differentiable. Stucki et al. introduced Betti matching, a method that pairs the persistence barcodes (the lifespans of topological features) of the predicted segmentation with those of the ground truth. By optimizing this match, the network learns not just to predict the right number of holes or components, but to ensure they exist in the exact correct spatial locations.

Similarly, Byrne et al. demonstrated that enforcing these persistent homology constraints prevents physically impossible errors, such as floating tissue islands or spurious voids in cardiac scans, bridging the gap between statistical pixel similarity and geometric validity.

Skeletonization and Centerline Alignment

Another highly intuitive strategy shifts the network's focus from the volumetric body of a structure to its "skeleton." By reducing a thick or thin tube down to a 1-pixel-wide centerline, any break in connectivity becomes glaringly obvious to the loss function.

Lv et al. leverage clDice (centerline Dice), a metric calculated by intersecting the skeleton of the prediction with the volume of the ground truth, and vice versa. Because a skeleton contains far fewer pixels than the full structure, a single disconnection causes a massive drop in the clDice score. This forces the network during backpropagation to prioritize maintaining continuous paths over perfectly refining boundary edges.

This topological focus is also essential when evaluating filamentous networks, where minor variations in width are clinically irrelevant compared to the integrity of the overall path length (Yi Liu et al.).

Geometric and Projection-Based Constraints

Beyond modifying loss functions, some researchers prevent fragmentation by changing how the data is represented or supervised.

Instead of classifying discrete voxels, the authors of some frameworks bypass the pixel grid entirely. For example, Xiaoyu Yang et al. represent branching coronary arteries as continuous geometric 3D meshes, which mathematically guarantees that the segmented output cannot break into disconnected pieces.

Alternatively, lower-dimensional projections can be used to safeguard higher-dimensional continuity. Zhanqiang Guo et al. demonstrated that 2D Maximum Intensity Projections (MIP) preserve the continuous, tree-like connectivity of sparse 3D vascular networks. By supervising 3D models with these continuous 2D "shadows," the network inherently learns global topological constraints that prevent localized fragmentation. Finally, Lai showed that models can even learn to self-correct these gaps by training on synthetic error maps derived from anatomical priors, actively filling in the topological breaks that standard architectures leave behind.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_topological_fragmentation_8cd3ab2a,
  title = {topological fragmentation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-fragmentation/rdr_8cd3ab2a3e5d4dcebf557d3d1c8c7883},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_topological_fragmentation_8cd3ab2a_v1,
  title = {topological fragmentation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-fragmentation/rdr_8cd3ab2a3e5d4dcebf557d3d1c8c7883/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake