Loading…
Cell type annotation from multiplexed tissue images, using a pretrained Vision Transformer ensemble that runs on new panels without fine-tuning.
No providers recorded yet. Browse all providers
Highly multiplexed tissue imaging — CODEX, MIBI, IMC — can measure more than 50 proteins in a single section, but turning those images into cell type calls is still mostly manual. Marker-range gating is reproducible only within a lab, supervised classifiers need a reference acquired with the same panel, and unsupervised clustering trails both. Nearly all also collapse each cell to average per-channel intensities, discarding the subcellular pattern in the image.
RIBCA, the Robust Image-Based Cell Annotator, comes from Huangqingbo Sun and Robert F. Murphy at Carnegie Mellon University with Emma Lundberg's group at Stanford University, KTH/SciLifeLab, Uppsala University and the Chan Zuckerberg Biohub. It classifies cropped single-cell images directly, and applies to a new image collection with no fine-tuning and no human in the loop.
Modularity is what makes this work. Rather than one model over one fixed marker set, RIBCA is an ensemble of Vision Transformers, each trained for a specific marker panel and the subset of cell types that panel can resolve. Given a user's metadata, the software determines which base models the available channels support, runs those, and merges their calls. Missing markers are filled in by dedicated channel-wise imputation models rather than being blanked.
Each base model is a small ViT, first pretrained by masked autoencoding on single-cell crops and then fine-tuned for classification with a label-smoothing cross-entropy loss, AdamW, and a learning rate of 2e-4. Both stages use SPData_1, a curated corpus of more than 15 million single-cell images assembled from 16 public multiplexed imaging datasets spanning CODEX and MIBI, with cell type names harmonized across datasets and a 25:1 training-validation partition. Preprocessing subtracts background by rolling-ball, clips each channel at the 98th percentile of nonzero pixels, rescales to [-1, 1], and applies a soft mask that blends the segmentation boundary so neighbouring signal fades rather than cuts off.
Generalization is tested on data acquired independently of SPData_1. The authors built SPHPAData_1 — 148 tissue microarray core images from 46 human tissues, imaged with a 46-plex panel and released in Human Protein Atlas v24 — and annotated over 1 million cells with RIBCA directly. Manual validation of selected cores gives roughly 82.2% fine-grained accuracy; the pixel-clustering method Pixie reaches 62.34% on the five cores where it assigned more than two or three cell types. On five public benchmark datasets, RIBCA's base ViT gives the best average and most stable cross-validation accuracy against CellSighter, MAPS and STELLAR, though no method wins on every dataset. Eighty-eight HuBMAP CODEX datasets were also annotated, with compositions agreeing with HPA for matched tissues. In total the tool has annotated over 3 million cells in more than 40 human tissues.
RIBCA targets spatial proteomics labs that have images and segmentation but no matched reference annotation — which is most of them. Needing neither training data nor per-experiment tuning, it converts cell typing from a bespoke analysis into a preprocessing step. The paper uses RIBCA calls to compute neighbourhood composition over 15-cell spatial niches and to partition lymph node images into subcapsular sinus, paracortex, germinal center and medulla.
RIBCA is peer-reviewed in Cell Systems, and the release includes GPL-3.0 inference code,
curation scripts for reassembling SPData_1, and a new annotated HPA tissue microarray dataset
useful as a benchmark in its own right. The weights are distributed less formally:
download_models.py pulls eight checkpoints — five base annotators, three imputation models —
from Google Drive with no license attached, and the Zenodo record archives the software, not the
weights. Neither training nor evaluation code ships, so the models cannot be retrained and the
benchmarks cannot be reproduced. The limits are stated in the paper: stromal cells
are the main systematic failure mode, often misclassified as epithelial, and the ensemble covers
only the cell types its five panels were built for, so anything outside them falls into an
"Others" bucket or must be added as a new base model. Training data is reproducible rather than
downloadable, and per-model parameter counts are not reported in the main text.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.