Supporting data for "Counting cells with confidence: uncertainty-aware atlas registration for whole-brain LSFM"

This record contains the data for "Counting cells with confidence: uncertainty-aware atlas registration for whole-brain LSFM". It holds everything needed to reproduce the substantia nigra pars compacta (SNpc) analyses from atlas registration onwards, for the five iDISCO-cleared, TH-immunolabelled mouse brains (Samples 1–5), together with the consensus regions of the aMAP expert-annotation validation. Each brain was registered to the Perens LSFM mouse atlas (20 µm) with brainreg, once for each of 48 combinations of bending-energy weight and grid spacing. This was done for two registration inputs: the raw autofluorescence and an intensity-preprocessed version. Registrations were filtered by global QC (tissue–atlas IoU > 0.9 and deformation grad_p95 < 0.045) and ranked by local midbrain SSIM, and the top half was retained. The retained registrations were combined into a voxel-wise inclusion-frequency map, and the consensus SNpc is the set of voxels with inclusion frequency ≥ 0.5. TH+ cells were segmented with a custom Cellpose model, and a cell is counted in a region if any of its voxels lies in it. What is included: Each brain has its own zip (Sample1.zip … Sample5.zip, 1.9–5.4 GB unzipped), with numbered folders that follow the pipeline: 01_registration_inputs/: the 20 µm autofluorescence (AF) images used for registration (raw and preprocessed), the 20 µm TH image and the tissue mask used for QC. Each is given both as passed to brainreg and in brainreg's coronal sample space, where it is voxel-aligned with every registration's SNpc mask and the 20 µm consensus maps, so the masks can be overlaid directly. 02_registrations/<raw|preprocessed>/: the SNpc mask from each of the 48 registrations, plus registrations.csv. That table has one row per registration: parameters, QC metrics (tissue–atlas IoU, grad_p95, midbrain SSIM), the SNpc TH+ count, and whether it passed global QC and was retained after SSIM ranking. 03_consensus_20um/<raw|preprocessed>/: inclusion-frequency maps and consensus masks at several thresholds, from the retained registrations and from all 48 registrations. 04_fullres_midbrain/<raw|preprocessed>/: the full-resolution TH crop around the midbrain, the Cellpose cell labels, the inclusion-frequency maps and 0.5 consensus mapped into that crop, the cells counted in the consensus, and the crop geometry. shared.zip (0.9 GB) contains: cellpose_model/: the trained Cellpose model for TH+ cells. tables/: the essential result tables: - QC metrics, selection and SNpc TH+ count of every registration; - consensus counts for every registration set and inclusion threshold; - the results of the two validation datasets: every Dice score of the aMAP expert-annotation validation, and the per-region comparison of the independent LSFM dataset with its published registration; - Supplementary Tables S1–S5. qc_sensitivity/: the global-QC threshold sensitivity analysis for all three datasets, including the full-resolution SNpc consensus masks, counted cells and agreement tables under each of the eight QC rules. aMAP_per_brain_galleries/: every aMAP structure in every brain, shown through the pipeline. aMAP_consensus_regions/: the consensus regions of the aMAP validation, so the comparison with expert anatomy can be checked without re-running the registrations. For each of the six brains, on one 25 µm grid: - the brain in brainreg's sample space and its tissue mask; - the inclusion-frequency map and 0.5 consensus for each of the 10 annotated structures; - the QC metrics and per-structure retention of all 64 registrations. The expert_STAPLE/ subfolder holds the expert STAPLE references (2-D coronal sections on the same grid), the annotated-hemisphere masks used for scoring, and a table of the scored planes. With these files, the published consensus Dice scores are reproduced exactly The images of the two validation datasets are not redistributed. They are available from their original sources (aMAP, Niedworok et al. 2016; ScienceDB, doi:10.57760/sciencedb.nb.00025). All other analyses and figure panels can be regenerated from these files with the code repository.