2026-09-17
3D-MAESTRO chains destriping, stitching, CCF registration and a ResNet detector; ~4,000 whole-brain datasets in two years, nuclear F1 0.88 vs Cellfinder 0.80 and Cellpose 0.72.
Whole-brain light-sheet microscopy already images centimeter-scale tissue at diffraction-limited resolution. A single specimen lands near a terabyte. The bottleneck is the software after acquisition: vignetting and stripe correction, tile stitching, registration to the Allen mouse Common Coordinate Framework (CCFv3.1), then fluorescent cell detection. Each step has a tool. Chaining them usually means a pile of lab-specific scripts that break when the machine changes. ClearMap and NuMorph run end-to-end but do not distribute, so they stall on terabyte volumes. BrainGlobe is modular but does not pin a reproducible workflow across compute environments. GPU-heavy segmentation is out of reach for a workstation.
This is an infrastructure gap. Another segmentation leaderboard will not close it.
Allen Institute Neural Dynamics and Brain Science built 3D-MAESTRO (3D-Microscopy Automation and Execution with Scalable Tools, Rendering, and Orchestration). Nextflow orchestrates versioned containers. Images sit in OME-Zarr; metadata follows the AIND schema. The same pipeline runs on cloud, on-prem HPC, and a local workstation. Every stage writes a stable interface, so one algorithm can be swapped and timed without touching the rest.
Six stages.
Artifact correction uses BaSiC for vignetting and a modified pystripe (wavelet plus FFT) for stripes, now reading cloud OME-Zarr. Stitching was decided on 1,473 datasets: BigStitcher was about twice as fast as TeraStitcher and more stable, so it stayed. Alignment quality is Inter-Image Cell Correspondence Error (ICCE), the distance between the same cell seen in overlapping tiles.
Registration is the design-heavy step. White matter is bright in SPIM autofluorescence and dark in the serial two-photon template that defines CCFv3.1. Direct intensity matching warps the wrong way. They built a 10 µm SPIM template (ST) from far-red autofluorescence of ten aqueously cleared brains, register each new sample to ST (same contrast), then apply a precomputed ST-to-STPT warp. ANTs does the deformable work, mixing mutual information on gradients with hand-drawn labels on small nuclei and large divisions.
Cell detection is two-stage instance detection. A GPU 3D Laplacian of Gaussian proposes spots with a high-recall, low-precision setting. An 18-layer 3D ResNet then classifies each proposal, taking both the signal channel and the autofluorescence background on a (14, 14, 26, 2) crop at 2x downsample. Training used 59,721 labels (23,399 cells, 36,322 false hits) from 30 mice. The decision threshold is the valley in that dataset's likelihood histogram. Intermediate volumes open in Neuroglancer.
About 4,000 whole-brain datasets ran through the pipeline in two years, with turnaround under a day. For a typical 1 TB, three-channel, 1.8 x 1.8 x 2 µm volume, Figure 1 reports roughly 10 h correction, 1.33 h alignment, 2.74 h fusion, 1.6 h registration, 6 h cell detection, 0.86 h quantification. Table 4 prices six 0.7-0.9 TB brains on AWS on-demand (January 2026): 25.82 ± 2.53 wall-clock hours and 47.22 ± 3.31 USD of compute, storage excluded. Channel-parallel destriping is the expensive step, about 22 USD.
Stitching: median ICCE fell from 12.6 µm (range 7.4-17.2) to 2.2 µm (1.6-3.0), a 5.7-fold drop. The 3D voxel diagonal is 3.24 µm, so residual error is about one voxel. Overlaps did not double-count cells.
Registration: five test brains with seven hand-drawn structures (LGd, RT, and others). Routing through ST beat direct STPT registration on Dice and surface distance (rank-sum p≤0.001). Per-structure Dice values live in the figure, not a table. Inter-rater Dice was 0.93; mean surface distance 12 µm. Those test brains also entered template construction. A second template built only from wild-type brains pointed the same way.
Cell detection on nine held-out nuclear-label blocks:
| Method | F1 | Runtime on sub-volume |
| 3D-MAESTRO detector | 0.88 | 13.90 ± 14.06 s |
| Cellfinder | 0.80 | 42.81 ± 11.49 s |
| Cellpose | 0.72 | 84.12 ± 35.73 s |
Hardware: 16 CPUs, one V100, 64 GB RAM. After fine-tuning on 7,500 cytosolic tdTomato labels, 18 pons datasets scored precision 0.91 ± 0.043, recall 0.74 ± 0.21, F1 0.80 ± 0.15. Recall is noisy.
The rest of the paper is application: enhancer AAV and transgenic whole-brain maps, LC retrograde topography, hippocampal anterograde spread, TRAP after water restriction, Neuropixels track reconstruction, and cingulate versus motor cortex input maps. Those are demos, not algorithm comparisons.
Labs that already run whole-organ SPIM get a reproducible cloud-and-cluster workflow, not a new segmentation paper. The modularity is not theoretical: stitching already moved from TeraStitcher to BigStitcher, and the cell classifier has been revised in production. A modality-specific template is the standard answer to contrast mismatch. The detector's practical gain is speed: about 6x faster than Cellpose on the same V100, with a higher F1.
Budget is tens of dollars of compute per brain plus a Nextflow-capable environment. The template and classifier are tied to aqueously cleared SmartSPIM mouse brains. A different clearing protocol or organ needs a new template and new labels. The authors plan more histology templates and a look at 3D foundation models.
The structures used to score registration also helped build the template-to-template warp. Leave-one-out and a wild-type-only template soften that, but the evaluation brains still sat in the ST construction set. Dice numbers are figure-only.
Detection reports centroids, not morphology. Cytosolic recall of 0.74 with a 0.21 standard deviation will miss dense or dim clusters. The Cellfinder comparison reused the authors' ResNet and only swapped the proposal stage, so it is not a full off-the-shelf Cellfinder run. Cellpose was fine-tuned on in-house data, which is fairer.
The 25.8 h Table 4 total matches "under a day" only because channels run in parallel; a serial single-channel job is longer. The 47 USD figure ignores storage. Code is public. The paper does not show an outside lab running it unmodified.