CROCHET: a versatile pipeline for automated analysis and visual atlas creation from single-cell spatialomic data

CROCHET is a versatile, open-source, end-to-end pipeline designed to democratize spatial omics by enabling the automated construction of large-scale, interactive spatially resolved cell atlases from complex single-cell data across diverse sample cohorts.

The Gist

Imagine you have a massive, incredibly detailed map of a bustling city, but instead of streets and buildings, this map shows every single person (cell) in a tissue sample, what they are wearing (proteins), and who they are talking to (interactions). This is what spatial biology does for our bodies. It's like having a super-high-definition photo of a tumor or a healthy organ, showing exactly where every cell is and what it's doing.

But here's the problem: These photos are so huge and complex that they are impossible for a human to analyze by hand. It's like trying to count every grain of sand on a beach while also figuring out which grains are talking to each other.

Enter CROCHET.

Think of CROCHET not just as a tool, but as a super-smart, automated construction crew that takes these messy, raw photos and turns them into a clear, interactive, 3D atlas. The name stands for "ChaRacterization Of Cellular HEterogeneity in Tissues," but you can think of it as the "Master Architect" for cellular cities.

Here is how CROCHET works, broken down into three simple steps:

1. The Cleanup Crew (Meta-module 1 & 2)

Before you can build a house, you have to clear the construction site. CROCHET starts by taking the raw, blurry, or messy images from the microscope and cleaning them up.

• Focus & Alignment: Imagine taking a stack of slightly blurry photos and automatically picking the sharpest one, then lining them up perfectly so the buildings (cells) don't look like they are sliding around.
• The "Ghost" Hunter: Sometimes, when scientists take these photos, old colors from previous steps don't wash away completely. It's like trying to paint a new wall over a sticky, old poster. CROCHET has a special algorithm that spots these "ghost" signals (residual fluorescence) and wipes them away so they don't trick the computer.
• The "Fake" Detector: Sometimes, the camera gets confused and sees a speck of dust as a cell. CROCHET learns to spot these fake signals (non-specific binding) and ignores them, ensuring we only count real cells.

2. The City Planners (Meta-module 3)

Once the site is clean, CROCHET starts organizing the data.

• The ID Badge System: It looks at what "clothes" (proteins) each cell is wearing to figure out who they are. Is this a soldier cell (immune cell)? A builder cell (epithelial cell)? It creates a digital ID badge for every single cell in the image.
• The "Who's Next Door?" Map: This is the magic part. CROCHET doesn't just count cells; it measures proximity. It asks: "How close is a soldier cell to a cancer cell?"
  • The Analogy: Imagine a crowded party. Most tools just count how many people are in the room. CROCHET, however, calculates exactly who is standing next to whom, how loud they are talking, and if they are forming small groups. It creates a "Social Network Map" of the tissue.
• The "Immunoprint": This is a special report card. If you are looking for a specific interaction (like a lock and key between a cancer cell and an immune cell), CROCHET creates a visual map showing exactly where these interactions are happening across the whole tissue. It's like a heat map showing where the "handshakes" between cells are happening.

3. The 3D Time-Traveler

Tissues are 3D objects, but microscopes usually take 2D slices (like slicing a loaf of bread). CROCHET can take two slices of bread that are right next to each other and stitch them together to rebuild the loaf in 3D. It aligns the cells from one slice with the slice next to it, creating a continuous, three-dimensional model of the tissue.

Why is this a big deal?

Before CROCHET, analyzing these complex images was like trying to solve a giant jigsaw puzzle in the dark, with pieces from different boxes mixed together. You needed a PhD in math and coding just to get started.

CROCHET is open-source and user-friendly. It's like giving everyone a pair of night-vision goggles and a robot assistant that sorts the puzzle pieces for you.

• For Doctors: It helps find exactly where immune cells are failing to attack a tumor, which could lead to better, more personalized cancer treatments.
• For Scientists: It allows them to compare different patients' tissues easily, finding patterns that were previously invisible.

In short: CROCHET takes the chaotic, noisy, and overwhelming data of modern biology and turns it into a clear, interactive, 3D storybook of how our cells live, talk, and fight together. It democratizes this technology, making it accessible to anyone who wants to understand the "city" inside our bodies.

Technical Summary

1. Problem Statement

Spatial biology technologies (e.g., cyclic immunofluorescence, CODEX, 10X Xenium) generate high-resolution data linking tissue composition to function. However, the field lacks robust, standardized analytical methods to handle the complexity of this data. Key challenges include:

• Data Heterogeneity: Difficulty in integrating diverse data modalities and image formats.
• Preprocessing Artifacts: Issues such as residual fluorescence (bleed-through) from previous staining cycles, non-specific binding (NSB), tissue loss/damage during cyclic imaging, and inconsistent focus across Z-planes.
• Quantification Limitations: Existing spatial metrics often rely on simple point-density calculations that fail to account for varying tissue architectures or expression levels, making cross-sample comparisons difficult.
• Accessibility: A lack of user-friendly, end-to-end, open-source tools that democratize spatial omics analysis for non-experts.

2. Methodology: The CROCHET Pipeline

CROCHET (ChaRacterization Of Cellular HEterogeneity in Tissues) is an end-to-end, modular, open-source pipeline designed to transform raw multiplexed tissue images into spatially resolved single-cell atlases. It is organized into three meta-modules:

Meta-module 1: Image Processing, Segmentation, and Feature Extraction

• Standardization: Converts hundreds of image formats (e.g., CZI) to standardized TIFF using OME Bio-Formats.
• Focus Evaluation: Automatically selects the sharpest Z-sections using a normalized variance-based statistical algorithm.
• Registration & Segmentation:
  • Aligns multi-cycle images by superimposing identical nuclei.
  • Segments nuclei using Mask R-CNN (trained on CycIFAAP data) and cell boundaries using membrane markers or nucleus expansion (default 3x nuclear size). Cellpose is also supported as an alternative.
• Feature Extraction: Extracts subcellular protein readouts (nucleus, membrane, cytoplasm), spatial coordinates, and morphological descriptors (circularity, Haralick features).
• TMA Core Detection: Uses HDBSCAN clustering to identify individual tissue cores within Tissue Microarrays (TMAs).

Meta-module 2: Quality Control (QC) and Data Cleaning

• Tissue Loss Detection: Identifies lost cells/tissue across imaging cycles by comparing nuclear staining intensity (Hoechst) against the first cycle. Cells with significant intensity drops are flagged and removed.
• Bleach Correction: Estimates and removes residual fluorescence (carry-over signals) from previous cycles using average signal intensities from bleached images, reducing the need for harsh chemical bleaching.
• Non-Specific Binding (NSB) Removal: Detects artificial signals that recur with high amplitude in identical patterns across consecutive cycles (regardless of antibody specificity) and removes them at the subcellular resolution level.
• Normalization: Adjusts for variations in light intensity and exposure times across cycles.

Meta-module 3: Downstream Analysis and Visualization

• Interactive Visualization: Built on Napari, allowing users to overlay cell masks, expression levels, and annotations on tissue images.
• Cell Typing: An interactive Flask-based application uses a hierarchical gating strategy to assign cell identities based on user-defined marker thresholds.
• Spatial Enrichment Score (SES): A novel metric that quantifies spatial relationships.
  • Unlike traditional density metrics, SES uses a Radial Distribution Function (RDF) weighted by expression levels.
  • It normalizes spatial proximity to local tissue density, enabling direct comparison of biomarker distributions across heterogeneous tissues and rare cell types.
  • Available at both single-cell and sample levels.
• Immunoprint: A framework to map receptor-ligand interactions (e.g., PD1-PDL1) across patient cohorts, visualizing spatial co-distributions similar to oncogenic "oncoprints."
• Neighborhood Analysis: Uses Louvain/Leiden clustering on spatial enrichment scores to identify tissue niches and intratissue communities based on cell-cell interaction patterns.
• 3D Reconstruction: Aligns adjacent tissue slices using Euclidean transformations based on cell point densities to reconstruct 3D tissue continuity.

3. Key Contributions

1. End-to-End Automation: A unified pipeline handling the entire workflow from raw image processing to 3D atlas construction, specifically optimized for cyclic imaging but applicable to other spatial transcriptomics platforms.
2. Novel Computational Metrics:
   • Spatial Enrichment Score: A density-normalized metric that overcomes the limitations of simple point-density calculations, allowing for robust cross-tissue comparison.
   • Immunoprint: A new visualization and quantification tool for spatial receptor-ligand interactions.
3. Advanced QC Algorithms: Automated detection and correction of tissue loss, residual fluorescence, and non-specific binding, which are critical for the integrity of cyclic imaging data.
4. Modularity and Accessibility: The pipeline is open-source, modular (allowing integration of new segmentation or analysis methods), and features an interactive UI for non-bioinformaticians.
5. 3D Capability: Unique capability to reconstruct 3D tissue maps from 2D adjacent slices.

4. Results and Validation

• Data Source: The pipeline was demonstrated using a cyclic immunofluorescence (CyCIF) dataset of 10 cycles and 5 channels from gastrointestinal tumor TMAs (derived from Dereli et al.).
• Performance:
  • Achieved 96.9% accuracy and a 0.83 Dice score in nucleus segmentation using the pre-trained Mask R-CNN model.
  • Successfully identified and removed lost tissue and non-specific binding artifacts, improving data completeness.
  • Generated spatial maps showing distinct cell neighborhoods and immunoprints of PD1-PDL1 interactions across 5 bowel adenocarcinoma patients.
• Comparative Advantage: The paper notes that CROCHET outperforms or complements existing tools (MCMICRO, SIMPLI, TRACERx-PHLEX) by offering a more comprehensive, automated, and user-friendly workflow specifically tailored for the nuances of cyclic imaging and large-scale cohort analysis.

5. Significance

• Democratization of Spatial Omics: By providing a user-friendly, interactive, and open-source tool, CROCHET lowers the barrier to entry for biologists and clinicians to analyze complex spatial data.
• Standardization: It establishes a standardized framework for preprocessing and analyzing spatial omics data, crucial for reproducibility in large-scale studies.
• Clinical Translation: The ability to generate "immunoprints" and map specific immune checkpoint interactions across patient cohorts offers a direct pathway to identifying precision immunotherapy targets.
• AI Readiness: The generation of comparable, normalized spatial metrics makes the data suitable for training Artificial Intelligence models for spatial biology, a critical step toward clinical AI applications.

In summary, CROCHET addresses the critical bottleneck in spatial biology analysis by providing a robust, automated, and visually intuitive pipeline that transforms raw, noisy multiplexed imaging data into biologically meaningful, 3D-resolved cellular atlases.