Pipelines

Every analysis step you run from the slide detail page maps to one of these pipelines. This reference lists what each one takes as input and produces as output.

Ingestion and viewing

Pipeline	Input	Outputs
Generate tiles	Uploaded slide image	Tile manifest, thumbnail, pyramid levels

You can set the tile size (default 256 px).

Tissue analysis

Pipeline	Input	Outputs
Detect tissue	Slide	Tissue mask, region bounding boxes, focus QC summary
Segment cells	Tiled slide	Cell overlay, cell count, confidence summary
Color deconvolution	Tiled H&E or IHC (H-DAB) slide	Per-stain grayscale images, per-stain mean intensity, DAB-positive area fraction for IHC
Train object classifier	Segmented cells + class-labeled annotation polygons	Random-forest cell classifier with cross-validated accuracy, per-class metrics, feature importances, and a predicted class + probability for every cell

Segmentation supports a threshold_baseline connected-components model or a watershed model that splits touching nuclei via distance-transform watershed; Cellpose/StarDist-based deep-learning detection is planned but not yet integrated.

Spatial quantification

Pipeline	Input	Outputs
Spatial quantification	Completed segmentation run	Density heatmap, nearest-neighbor stats, region metrics, infiltration phenotype, CSV/JSON exports
Multiplex IF phenotyping	Completed segmentation run + multi-channel marker image	Per-cell marker positivity, phenotype calls (e.g. CD8+ T cell, CD68+ macrophage, PD-L1+, or a custom marker-combination rule), co-expression matrix, region composition, phenotype-map overlay

Both require Segment cells to complete first.

Whole-slide and TMA utilities

Pipeline	Input	Outputs
TMA dearray	TMA slide	Detected core centroids quantized onto a grid, row/column labels (A1, B2, …), missing-core flags
Export OME-TIFF	Processed slide	Pyramidal, tiled OME-TIFF with physical pixel size metadata, for opening natively in QuPath, Fiji/ImageJ, or OMERO

Cross-area spatial alignment

Pipeline	Input	Outputs
Spatial align	Pathology slide + a completed Computational Biology spatial run or dataset	Landmark or automatic affine coordinate transform, spot/cell overlay points on the pathology slide, region-level expression/cluster-composition summary

Consumes a compbio spatial run (or its .h5ad dataset directly) rather than duplicating spatial transcriptomics ingestion — see Spatial omics context.

Cohort

Pipeline	Input	Outputs
Cohort comparison	Multiple slides with sample group labels	Group comparison of pathology biomarkers: t-test/Mann-Whitney (two groups) or ANOVA/Kruskal-Wallis (three or more), Benjamini-Hochberg FDR, effect sizes and confidence intervals

Available as a full workflow on the study Cohort page: CSV sample import, per-slide sample/group/timepoint assignment, a batch pipeline launcher across the cohort, and the comparison table itself.

Interpretation

Pipeline	Input	Outputs
Interpret	One or more completed pathology runs	Grounded narrative citing computed metrics: key findings, hypotheses, caveats, research-use disclaimer

Open at /areas/pathology/studies/{id}/interpret. This page is not yet linked from the study navigation — reach it via the direct URL.

Job lifecycle

Every run moves through the same states, shown as a status badge on the slide detail page:

queued → running → complete | failed

Versioning

Every run records a pipeline version. Include run IDs and versions in methods sections for reproducibility.

Implementation sources

Tissue detection and baseline segmentation use scikit-image and OpenCV-based methods. Color deconvolution follows Ruifrok & Johnston (2001) stain-vector separation, in the style of QuPath/Fiji stain separation. TMA dearraying and the object classifier follow QuPath's TMA dearrayer and trainable object-classifier conventions. Cellpose, StarDist, and Macenko/Reinhard stain normalization remain future work.