TopROI: A topology-informed network approach for tissue partitioning

TL;DR

TopROI addresses the ROI-definition bottleneck in large tissue point clouds by marrying geometry-based proximity networks with higher-order topology from persistent homology. It builds a fused network G=(V,E,W) that combines a Delaunay-derived geometric layer with a topological layer formed from persistent cycle representatives, and then partitions this network into biologically meaningful ROIs using Leiden optimization of the Reichardt–Bornholdt Potts model with a tunable resolution γ. In synthetic gland-like data, TopROI more faithfully preserves glandular structures and yields persistence diagrams that closely match ground truth; in human colorectal tissue, ROI topologies reveal a continuum of architectural disruption from healthy mucosa to carcinoma, enabling ROI-level topological analyses that capture intra- and inter-sample heterogeneity. The approach provides a principled, modular framework for analyzing tissue organization across scales, with potential applicability to other structured tissues and disease contexts where architecture is a diagnostic hallmark.

Abstract

Mammalian tissue architecture is central to biological function, and its disruption is a hallmark of disease. Medical imaging techniques can generate large point cloud datasets that capture changes in the cellular composition of such tissues with disease progression. However, regions of interest (ROIs) are usually defined by quadrat-based methods that ignore intrinsic structure and risk fragmenting meaningful features. Here, we introduce TopROI, a topology-informed, network-based method for partitioning point clouds into ROIs that preserves both local geometry and higher-order architecture. TopROI integrates geometry-informed networks with persistent homology, combining cell neighbourhoods and multiscale cycles to guide community detection. Applied to synthetic point clouds that mimic glandular structure, TopROI outperforms quadrat-based and purely geometric partitions by maintaining biologically plausible ROI geometry and better preserving ground-truth structures. Applied to cellular point clouds obtained from human colorectal cancer biopsies, TopROI generates ROIs that preserve crypt-like structures and enable persistent homology analysis of individual regions. This study reveals a continuum of architectural changes from healthy mucosa to carcinoma, reflecting progressive disorganisation in tissue structure. TopROI thus provides a principled and flexible framework for defining biologically meaningful ROIs in large point clouds, enabling more accurate quantification of tissue organization and new insights into structural changes associated with disease progression.

Figures (11)

• Figure 1: Overview of TopROI pipeline. Top: construction of a Delaunay-based spatial graph, commonly used to represent cell contact networks; centre: full pipeline of TopROI; bottom: persistent homology extracts topological information from point clouds, in our case physical representatives of data features.
• Figure 2: TopROI preserves geometry and topology in synthetic data. (a) Partition of synthetic dataset into 16 ROIs (black Voronoi borders are not a result of the algorithm, and are included for visualisation only); (b) ellipse markers example, and ellipse preservation results; (c) comparison of original and reconstructed 1D persistence diagrams with noise band (red), and average Wasserstein distances; (d) Example alpha shapes of TopROI ROIs, and average area variation, circularity ($\pi/4$ for a square, dashed) and convexity (1 for a square, dashed).
• Figure 3: Topological analysis and clustering of ROIs from biopsied tissue samples. (a) Topological analysis pipeline. (b) 2D PCA projection of vectorised features. (c) K-means clustering results projected onto PCA space, and minimum spanning tree connecting cluster centroids (black line); cluster IDs follow tree order. (d) Disease stage enrichment scores for each topological cluster. (e) Topological profiles of individual samples, grouped via hierarchical clustering (Groups, brown shades). Groups are obtained with the default parameters of scipy.cluster.hierarchy2020SciPy-NMeth.
• Figure 4: Results of (a) inertia and (b) silhouette scores from 50 $k$-means runs for different numbers of clusters. The elbow at $k=7$ is indicated by a dashed line.
• Figure 5: Images of the 15 human colorectal tissue biopsies (3 healthy (H), 6 benign lesions (B), 6 cancerous (C)), each divided into regions of interest using TopROI. Regions are coloured according to the $k$-means clustering of their vectorised topological features (see Section \ref{['subsec:Results_real']}).