Application of the presented framework to 3D+t light-sheet microscopy images of zebrafish embryos

The panels show the maximum intensity projections on the XY, YZ and XZ planes. Seed points were extracted using the LoGNSM+F+U method and the improved seed points were used for segmentation using the TWANG+U method. Extracted objects of two opposite views were combined using a segment-based fusion approach (S4 Note) and the fused object locations were then used for tracking (TWANG+U+NN). The first three columns show results of a single view at approximately 8 hours post fertilization (hpf) with the color-code according to the FSMD values of the extracted objects (red over blue to green for low, medium and high FSMD values). The last column shows the results after multiview fusion and tracking at about 10 hpf using a cell density for coloring. In addition, the embryo was oriented along the coordinate axes, such that the anteroposterior axis was aligned with the Y axis (animal pole on the top) and the dorsoventral axis aligned with the X axis (dorsal to the right) [37]. Data were taken from our previously published work [37–39, 47]. Scale bar, 100 μm.

Exemplary comparison of applying OTSU and TWANG to a large-scale light-sheet microscopy image of fluorescently labeled cell nuclei of a zebrafish embryo(single time point at about 9 hours post fetilization).

OTSU completely failed to extract meaningful information in this case and only captured a few correct nuclei by chance. TWANG could resolve most of the nuclei correctly and by design largely avoids under-segmentation errors. As the results provided by the plain OTSU algorithm were not properly extracting the embryonic shape, the watershed-based post-processing steps used in OTSUWW and OTSUWW+U were skipped. The results of TWANG+U are visually indiscernible from the plain TWANG results and we thus only show one panel for convenience. Furthermore, due to the image size of about 5 GB per frame, applying the 3D watershed algorithm on the entire image was not possible given the memory limitations of 32 GB on our test computer. A lack of ground truth for these large-scale data sets only allowed a qualitative comparison (see [39] for details). Scale bar: 100μm.