This repository holds the imaging mass cytometry (IMC) analysis of the chondrosarcoma (CHS) tissue microarray (TMA) of the Humanitas sarcoma cohort. The aim is to describe the tumour microenvironment of chondrosarcoma at single-cell level: which immune and stromal cells are present, how many, and where they lie relative to the tumour.
The workflow follows the Bodenmiller IMC pipeline (Windhager et al., 2023) and adds the steps that sarcoma TMAs need. A pathologist reviews the data at four points: the H&E section of every core, the annotation of artifacts and of the evaluable area, the segmentation, and the choice of markers for phenotyping.
The same workflow was later applied to the chordoma TMA (imc_analysis_chordoma).
raw channel images (imaging facility export)
-> panel generation
-> hot-pixel filter, channel stacks
-> Cellpose segmentation (nuclear + membrane composites)
-> intensity, morphology and neighbour measurement
-> CSV / AnnData / GraphML export
-> pathologist's manual review: evaluable area and excluded areas per core
-> cell status per object (clean, counted only, removed)
-> DNA status per core
-> core information and cell selections joined to every cell
-> large-object review
- Every step is a numbered script in
scripts/, called from a notebook cell and followed by a validation cell. - Each script writes a log to
logs/and numbered outputs toresults/. - The pathologist's decisions are stored as tables in
data/manual_review/and applied by script; they are never edited by hand downstream. - Flagged cores and cells are kept, so that their effect can be tested in sensitivity analyses.
| Notebook | Role |
|---|---|
notebooks/01_Complete IMC Data Analysis Workflow.ipynb |
Raw inputs, panel, channel stacks, Cellpose segmentation, measurement, neighbours and export, each with a validation cell |
The TMA holds cores from 15 patients, 11 positions per patient, sampled from defined areas of each tumour.
| TMA row | Sampled area |
|---|---|
| A–C | High-grade tumour |
| D | Low-grade tumour |
| E–F | Tumour with high inflammatory infiltrate |
| G–H | Tumour periphery, inside the tumour edge |
| I–J | Tumour border, non-neoplastic tissue just outside the edge |
| K | Non-neoplastic tissue far from the tumour |
The patient number equals 16 minus the TMA column. ROIs are named by
acquisition number and TMA position, for example ROI101_C3.
| Item | Value |
|---|---|
| Positions on the TMA | 165 |
| ROIs acquired (Hyperion, 1 µm) | 136 |
| Channels per ROI (35 antibodies, 2 DNA, 6 technical) | 43 |
| Cores included after the H&E review | 121 |
| Cores excluded: content did not match the sampled area | 8 |
| Cores excluded by design: non-neoplastic row K | 7 |
| Path | Purpose |
|---|---|
data/panel.csv |
Steinbock panel: channels, names, segmentation roles |
data/roi_marker_availability.csv |
Channels present in each ROI |
data/images.csv |
Image metadata |
data/cells.csv, data/cells.h5ad |
Single-cell table and AnnData object |
data/manual_review/core_areas.csv |
Per core: evaluable area, excluded area, density denominator, DNA status, matrix %, inclusion |
data/manual_review/cell_flags.csv |
Per object: position relative to the evaluable and excluded areas, cell_status |
docs/review/ |
Pathologist's review of the pipeline |
docs/presentations/ |
Slides on the workflow and the manual review |
notebooks/ |
Documented workflow |
scripts/ |
Numbered scripts; scripts/r/ holds R helpers for thresholds and clustering checks |
logs/ |
Run logs |
results/ |
Numbered tables and figures |
The channel stacks (about 10 GB), masks and per-cell measurement folders are not stored in the repository; they are rebuilt by steps 2–5. The output folders of scripts 11–13 are also rebuilt by running the scripts.
| Role | Channels | cellpose value |
|---|---|---|
| Nuclear | DNA1 (191Ir), DNA2 (193Ir), SOX9 | 1 |
| Membrane / cytoplasm | S100, CD68, aSMA, CD31, CD3, CD66b | 2 |
| Technical, not analysed | 127I, 131Xe, 134Xe, 138Ba, 208Pb, 80ArAr | – |
SOX9, a nuclear marker of chondrosarcoma cells, is added to the nuclear image because the DNA signal is weak or absent in many matrix-rich cores. CD20 is left out for lack of signal and CD138 for its non-specific signal.
Script: scripts/01_create_panel_from_raw.py
- Build the panel from the channel files present in the raw data.
- Assign the nuclear and membrane segmentation markers.
- Flag technical channels with
analysis = 0.
Script: scripts/02_preprocess_imc_images.py
- Apply the steinbock hot-pixel filter to every channel: a pixel more than 50 counts above all eight neighbours is set to its brightest neighbour.
- Stack the channels into one TIFF per ROI, named after the TMA core map.
scripts/validate_ome_tiff.pychecks that every stack is a valid OME-TIFF whose channel names match the panel.
Script: scripts/03_segment_cellpose.py
- Build a nuclear and a membrane composite, each the pixel-wise maximum of its markers scaled between the 1st and 99.5th percentiles.
- Run Cellpose 4.2 (
cpsam_v2) without retraining; model, versions and parameters are logged.
Scripts: scripts/04_measure_intensities.py, scripts/05_measure_regionprops.py,
scripts/06_measure_neighbors.py, scripts/07_export_data.py
- Mean intensity of every channel over the pixels of each cell.
- Area, centroid, axis lengths and eccentricity.
- Neighbours: cells whose outlines, expanded by 4 px, touch.
- Export as
cells.csv,cells.h5adand one GraphML graph per ROI.
Script: scripts/08_corner_background_correction.py
- Derive a tissue mask from the marker stack and set pixels outside the core to zero, without changing intensities inside the tissue.
Scripts: scripts/10_apply_manual_review.py, scripts/apply_manual_review.py,
scripts/09_masks_to_geojson.py
Outputs: data/manual_review/cell_flags.csv, data/manual_review/core_areas.csv
The pathologist drew the evaluable area of every core in QuPath and annotated barium and lead deposits, antibody aggregates, single-antibody spots, folds and necrosis. For every object:
cell_status |
Rule | Use |
|---|---|---|
outside_evaluable_area |
Centroid outside the evaluable area | Removed |
clean |
No pixel in an excluded area | Densities and phenotyping |
phenotype_not_assessable |
Up to 75% of pixels in excluded areas | Densities only |
aggregate_removed |
More than 75% of pixels in excluded areas | Removed |
The density denominator of each core is the evaluable area minus the excluded areas. Script 09 writes the cell outlines, coloured by status, as QuPath GeoJSON for visual checks.
Script: scripts/11_dna_status_per_core.py
- An object is DNA-negative when arcsinh(DNA1) + arcsinh(DNA2) < 1.
- A core is valid below 20% DNA-negative objects, intermediate at 20–50% and DNA-negative above 50%.
Script: scripts/12_annotate_cells.py
- Join to every cell the core information (patient, sampled area, DNA status, inclusion, H&E flags, matrix %) and its review flags.
- Define the selections used downstream: included cores, cells counted for densities, cells available for phenotyping.
Script: scripts/13_large_object_review.py
- Apply the pathologist's classification of objects larger than 400 px: real large cells are kept, non-cellular objects removed, uncertain ones flagged for sensitivity analysis.
| Output | Description |
|---|---|
results/01_panel.csv |
Panel copy |
results/01_roi_marker_availability.csv |
Channels per ROI |
results/01b_channel_preview_ROI001_E10.png |
Channel preview of one core |
results/02_images.csv |
Image metadata copy |
results/03_preprocessed_tiff_inventory.csv |
TIFF stack inventory |
results/04_cellpose_mask_inventory.csv |
Mask inventory |
results/05_intensity_table_inventory.csv |
Intensity table inventory |
results/06_regionprops_table_inventory.csv |
Region property inventory |
results/07_neighbors_table_inventory.csv |
Neighbour table inventory |
results/08_cells.csv |
Exported single-cell table |
results/09_cells.h5ad |
Exported AnnData object |
results/10_graphml/ |
One spatial graph per ROI |
results/hot_pixel_check/ |
Before/after check of the hot-pixel filter on one core |
| Stage | Status |
|---|---|
| Panel generation | Complete |
| Hot-pixel filter and channel stacks | Complete |
| Cellpose segmentation | Complete |
| Single-cell measurement and export | Complete |
| H&E review and core selection | Complete |
| Artifact and evaluable-area annotation | Complete |
| Segmentation review | Complete |
| Marker selection for phenotyping | Complete |
| Large-object review | Complete |
| Cell phenotyping | In progress |
| Comparison with the IHC counts of the same cores | Planned |
| Spatial analysis | Planned |
| Metric | Value |
|---|---|
| Segmented objects, 136 ROIs | 79,739 |
| Objects in the 121 included cores | 62,792 |
| Clean cells, available for phenotyping | 56,734 |
| Counted only (partly on an excluded area) | 3,516 |
| Removed as aggregates | 1,880 |
| Centroid outside the evaluable area | 662 |
| Included cores with adequate / reduced / poor DNA signal | 36 / 18 / 67 |
| Median chondroid matrix on H&E, adequate vs poor DNA signal | 10% vs 80% |
| Markers suitable for phenotyping | 16 of 35 |
- The DNA signal is poor in 67 of 121 cores, mostly those rich in chondroid matrix, although the nuclei are intact on H&E. In these cores some non-tumour cells without a bright membrane marker may not be segmented.
- No spillover matrix was acquired for this panel, so crosstalk between adjacent metal channels is not corrected.
- The segmentation was checked visually, not against manual outlines.
- Only 16 of 35 markers are suitable for phenotyping; checkpoint markers are not used.
Complete the phenotyping with the pathologist's gating rules and a supervised classifier, check the cell types against the IHC counts of the same cores, and then run the spatial and sensitivity analyses.
- Windhager J, Zanotelli VRT, Schulz D, et al. An end-to-end workflow for multiplexed image processing and analysis. Nat Protoc. 2023;18(11):3565–3613. https://doi.org/10.1038/s41596-023-00881-0
- Stringer C, Wang T, Michaelos M, Pachitariu M. Cellpose: a generalist algorithm for cellular segmentation. Nat Methods. 2021;18(1):100–106. https://doi.org/10.1038/s41592-020-01018-x
- Zanotelli VRT, Bodenmiller B. ImcSegmentationPipeline: a pixel classification based multiplexed image segmentation pipeline (v0.9). Zenodo; 2017. https://doi.org/10.5281/zenodo.3841961