The problem — A 3D culture sitting at the bottom of a well is not a flat cell. Treating it like one — projecting z-stacks, segmenting 2D max-intensity images, reading viability from a bulk fluorescence signal — discards precisely the volumetric and morphological information that makes the model system worth running. The mis-step is invisible until the dose-response curve disagrees with the orthogonal assay.
What it is / how it works — Organoids and spheroids are self-organizing 3D cultures that recapitulate tissue architecture far more faithfully than monolayers. For computational analysis the pipeline starts with z-stack acquisition on confocal or light-sheet systems, read by Bio-Formats into indexed per-well volumes. Preprocessing restores signal depth with RedLionfish or clij2 deconvolution; without it, structures more than ~30 µm below the coverslip are blurred into unreliable shapes. Segmentation runs Cellpose, StarDist, or MorphoLibJ in 3D — producing object masks from which scikit-image extracts volume, sphericity, surface roughness, and viability proxies (mean live-channel intensity, dead-channel fraction). Those features feed a dose-response or morphological-classification readout: growth curves, IC50 fits, class shifts that distinguish cytostatic from cytotoxic treatment.
High-content screening of organoids requires light-sheet microscopy for rapid z-coverage with low phototoxicity, combined with ML-based segmentation to handle the morphological diversity across wells, conditions, and timepoints. Recent reviews confirm that image-based profiling of size and shape outperforms bulk-viability readouts for detecting nuanced drug effects, including an 88 % positive predictive value for patient treatment response in one organoid platform (Lampart et al., 2023).
Where it breaks — The QC layer must catch two failure modes that do not appear in 2D pipelines. First, z-coverage: if the stack stops before the bottom of the organoid, volume and sphericity estimates are truncated and not obviously wrong. Second, debris and empty wells: small fragments trigger false segmentations that inflate object counts and skew morphology distributions. Both require per-well QC metrics (coverage fraction, object-area histogram, empty-well flag) before any feature is trusted — as discussed in Evaluating Segmentation Models in Real Microscopy Workflows. The same depth-dependent signal attenuation that makes Light-Sheet Microscopy — The Pipeline Problem Behind the Beautiful Volume challenging in whole-organ volumes reappears at smaller scale here; deconvolution is not optional.
Intensity integrates living and dead cells, debris, and background — all confounded by z-position. Segment in 3D first; then compute viability from per-object channel ratios within the mask. Skipping segmentation to save compute time reliably produces a noisier readout than the biology warrants.
How we run it — the Organoids / 3D pipeline, stage by stage — the nine-stage imaging backbone reshaped for this modality, with the tools, sub-steps and deliverable at each stage.