TopoTome: Topology-informed unsupervised segmentation and analysis of 3D images
Biological systems are three-dimensional and complex. Today, images of biological structures can be acquired using different imaging technologies and at increasing resolutions. However, identifying relevant structural features in three-dimensional (3D) images remains a significant challenge. 3D image segmentation is usually performed using deep-learning segmentation models. Such models are trained on manually annotated and segmented, dataset-specific images. Consequently, they rarely generalize across datasets. Here, we overcome these limitations with TopoTome, an unsupervised 3D image segmentation and analysis algorithm. TopoTome is based on topological data analysis, and conceptually distinct from standard clustering and deep learning image analysis models. It encodes the 3D image directly in topological space. Then, it performs unsupervised clustering to detect and segment features represented by salient and spatially coherent voxel intensity gradients. We demonstrate on simple, complex, synthetic and real-world 3D image data that it outperforms all 3D clustering algorithms. Its segmentation ranks with or outperforms best-in-class deep learning 3D image segmentation software. Beyond image segmentation, it provides streamlined topological data analysis of 3D images, advancing 3D image analysis from conventional mesh volumetry to structural topology. Owing to its conceptually different topological data analysis core, TopoTome does not need prior information and tuning to generalize across different imaging modalities, including fluorescence microscopy and X-ray computed tomography. We show it also readily generalizes across biological subjects, such as different species, organs and cells. TopoTome is thus one of the most versatile and accurate unsupervised 3D image segmentation algorithms.