A1 confocal microscope

Immunostained slices were imaged with a Nikon A1 confocal microscope controlled with the NIS-elements advanced imaging software or a Leica SP5 confocal microscope controlled by LAS AF. Images were collected at 4x, 10x, 20x, 40x, and 63x. To ensure consistency in imaging, the same confocal microscope was used within a data set and the parameters for offset, pinhole, gain and laser intensity were kept constant for each experimental set.

Mouse thyroids were quickly dissected and immersion-fixed in 10% formalin, paraffin-embedded, sectioned, and stained with hematoxylin and eosin (Vector Laboratories). For immunofluorescence, thyroid sections (6 μm) were deparaffinized in Citrisolv, then rehydrated using a graded ethanol series, followed by antigen retrieval in citrate buffer, blocking in 1.5% goat serum, incubation with primary antibodies (overnight, 4°C) and Alexa Fluor–conjugated secondary antibodies (Invitrogen A11073, A11001, A21422, A21428, A21245; Jackson ImmunoResearch 712-606-153; 1 hour, room temperature), counterstaining with ProLong Gold and DAPI (Invitrogen), and imaging with Nikon A1 confocal microscope or Leica STELLARIS 8 FALCON confocal microscope. Quantification of CD45⁺ cells in the proportion of total cells was performed using AIVIA Artificial Intelligence-guided Software. Immunohistochemistry of Ki67 used VECTASTAIN-ABC (Vector Laboratories) with 40× objective image capture (Olympus EX51 Microscope). Quantitation of Ki67-positive nuclei as a fraction of total thyroid nuclei per field, or Mac2-immunostained cells as a fraction of total thyroid nuclei per field, was performed using Imaris software (version 7.7.2).

The algorithm was tested on datasets acquired with a Nikon A1 confocal microscope (40x objective, excitation length 457 nm, bandwidth 500–550 nm), and a LEICA SP8 STED 3DX (93x objective, pulsed white-light laser 598 nm, pulsed 775 nm depletion laser, bandwidth 605–777 nm). All image stacks were from Purkinje cells within murine cerebella, cleared as in Magliaro et al.^{30 (link)}. We evaluated SENPAI’s performance in distinguishing both cellular (i.e., neurons) and subcellular (i.e., spines) structures.
The first 40x confocal image is a 512-by-512-by-143 image containing 114 somas (Fig. 2A), while the second one is a 512-by-512-by-139 image containing 103 somas. SENPAI was used on these datasets with a single clustering instance, without Gaussian smoothing. Then, post-processing routines including parcellation and pruning were exploited for isolating single cells.
The 93x STED dataset is a 1024-by-1024-by-35 image in which several sections of neurons can be observed (Fig. 2A). The segmentation with SENPAI was obtained merging two clustering instances with different levels of smoothness (i.e., setting the standard deviations of the 3D Gaussian filter to 0 and 3). Finally, we assigned the dendritic spines to their parent branch.