Instance Segmentation with Supervoxel Based Topological Loss Function

Anna Grim, Jayaram Chandrashekar, Karel Svoboda, Uygar Sümbül

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Instance segmentation, deep learning, digital topology

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose a supervoxel based topological loss function for training neural networks to perform instance segmentation.

Abstract: Reconstructing the intricate local morphology of neurons as well as their long-range projecting axons can address many connectivity related questions in neuroscience. While whole-brain imaging at single neuron resolution has recently become available with advances in light microscopy, segmenting multiple entangled neuronal arbors remains a challenging instance segmentation problem. Split and merge mistakes in automated tracings of neuronal branches can produce qualitatively different results and represent a bottleneck of reconstruction pipelines. Here, by extending the notion of simple points from digital topology to connected sets of voxels (i.e. supervoxels), we develop a topology-aware neural network based segmentation method with minimal overhead. We demonstrate the merit of our approach on a newly established public dataset that contains 3-d images of the mouse brain where multiple fluorescing neurons are visible as well as the DRIVE 2-d retinal fundus images benchmark.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8037