NucMM Dataset: 3D Neuronal Nuclei Instance Segmentation at Sub-Cubic Millimeter Scale

Segmenting 3D cell nuclei from microscopy image volumes is critical for biological and clinical analysis, enabling the study of cellular expression patterns lineages. However, current datasets neuronal usually contain smaller than $10^{\text{-}3}\ mm^3$ with fewer 500 instances per volume, unable to reveal complexity in large brain regions restrict investigation structures. In this paper, we have pushed task forward sub-cubic millimeter scale curated NucMM dataset two fully annotated volumes: one $0.1\ electron (EM) volume containing nearly entire zebrafish around 170,000 nuclei; $0.25\ micro-CT (uCT) part a mouse visual cortex about 7,000 nuclei. With imaging modalities significantly increased size instance numbers, discover great diversity appearance density, introducing new challenges field. We also perform statistical analysis illustrate those quantitatively. To tackle challenges, propose novel hybrid-representation learning model that combines merits foreground mask, contour map, signed distance transform produce high-quality masks. The benchmark comparisons on show our proposed method outperforms state-of-the-art segmentation approaches. Code data are available at https://connectomics-bazaar.github.io/proj/nucMM/index.html.