DeepThresh

Supervised deep-learning thresholding of imaging mass cytometry (IMC) channels.

DeepThresh replaces manual, per-channel intensity thresholding of IMC images with a bank of small segmentation networks: one binary U-Net per channel, trained on a handful of expert-thresholded regions of interest (ROIs) and then applied to every ROI in the dataset. It was developed to threshold extracellular-matrix (ECM) and cellular markers in mouse lung IMC data for the study below.

Parkinson JE, Bryant M, Ghafoor M, et al. Extracellular matrix phenotyping by imaging mass cytometry defines distinct cellular matrix environments associated with allergic airway inflammation. Molecular Systems Biology (2026). doi:10.1038/s44320-026-00234-5

How it works

Stage	Detail
Input	A single IMC channel (one metal / protein) from each ROI, as a 2-D TIFF
Normalisation	Every image is histogram-matched to a reference ROI, then min-max scaled to [0, 1]
Tiling	Overlapping 224 × 224 windows (stride 28) for training; non-overlapping zero-padded 224 × 224 blocks at inference
Model	segmentation_models_pytorch U-Net with a ResNet-152 encoder (ImageNet weights, frozen), 1 input channel, 1 output channel, sigmoid activation
Training	Dice loss, Adam (lr 1e-4), batch 32, 10 epochs; the checkpoint with the best validation IoU (threshold 0.5) is kept
Augmentation	Horizontal flip, shift / scale (scale ±0.5, shift ±0.1), random 224 × 224 crop (albumentations)
Output	Binary mask per channel per ROI, reassembled from the predicted blocks

One model is trained per channel, so a 37-channel panel yields 37 independent models. Each model only ever sees its own channel, which keeps the problem simple and lets a channel be re-trained on its own if the expert thresholds change.

Data

The paper's IMC dataset comprises 36 ROIs of mouse lung (BALB/c and C57BL/6 mice, allergen-challenged vs. PBS control) acquired with a 37-marker panel that includes ECM targets (collagens I / III / IV / VI, fibronectin, laminin-γ1, hyaluronan-binding protein, heparan and chondroitin sulphate, fibrinogen) alongside immune, epithelial and stromal markers.

Six ROIs were thresholded by hand by an expert biologist and used as the supervised signal for every channel:

Split	ROIs
Train	4
Validation	1
Test	1

Raw images and expert masks are not distributed in this repository. Trained weights are available on request (see Contact).

Repository layout

DeepThresh/
├── utils/
│   ├── model_per_channel_utils.ipynb   # preprocessing, tiling, dataset, training loop (train_single_models)
│   └── trained_model.ipynb             # single_ch_model: load a trained .pth and threshold new images
├── Results/
│   ├── In115_prediction.png            # example: Ym2 channel
│   └── Nd146_prediction.png            # example: SPC channel
└── docs/
    └── deepthresh_overview.svg

Installation

Python ≥ 3.8 with a CUDA-capable GPU is recommended. The training utilities rely on the smp.utils training loop that was removed from later releases of segmentation_models_pytorch, so the version is pinned:

pip install torch torchvision
pip install segmentation_models_pytorch==0.1.0 albumentations scikit-image tifffile scikit-learn pandas matplotlib tqdm

Usage

1. Train one model per channel

utils/model_per_channel_utils.ipynb expects two directory trees with matching file names per channel — raw channel TIFFs and the corresponding expert binary masks — and globs them with the channel identifier (e.g. Nd146). Adjust the two glob() paths at the top of preprocess_data() to point at your data, then:

targets = ['In115', 'Nd146', 'Tm169', ...]   # one entry per channel / metal tag
train_single_models(targets)

For each target this writes:

single_model_results_unet_v2/
├── trained_models/<target>_best_model.pth    # best-validation-IoU checkpoint
├── model_results/<target>_metrics.csv        # test-set accuracy / precision / recall / F1
└── test_predictions/<target>_prediction.png  # raw | manual threshold | predicted threshold

2. Threshold new images with a trained model

utils/trained_model.ipynb defines single_ch_model, which loads a checkpoint, normalises a list of images of that channel, predicts block-wise and stitches the blocks back to full-ROI binary masks:

from tifffile import imread

imgs = [imread(p) for p in nd146_paths]          # list of 2-D arrays, one per ROI
model = single_ch_model('trained_models/Nd146_best_model.pth', imgs)
model.predict()
masks = model.thresholded_images                 # (n_rois, H, W) binary array

Note. predict() currently reshapes each ROI into a fixed 4 × 3 grid of 224 × 224 blocks (the ROI size used in the paper). For ROIs of a different size, change the two reshape calls in predict() to match the block grid returned by image_to_blocks().

Example results

Raw channel, expert manual threshold and DeepThresh prediction on the held-out test ROI for two channels:

Citation

If you use DeepThresh, please cite the paper it was developed for:

@article{Parkinson2026_ECM_IMC,
  title   = {Extracellular matrix phenotyping by imaging mass cytometry defines distinct cellular matrix environments associated with allergic airway inflammation},
  author  = {Parkinson, James E. and Bryant, Morgan and Ghafoor, Mohamed and Dodd, Rebecca J. and Tompkins, Hannah E. and Fergie, Martin and Burgess, Matthew O. and Rattray, Magnus and Sutherland, Tara E.},
  journal = {Molecular Systems Biology},
  year    = {2026},
  doi     = {10.1038/s44320-026-00234-5},
  url     = {https://doi.org/10.1038/s44320-026-00234-5}
}

A machine-readable version is in CITATION.cff (GitHub's "Cite this repository" button).

Related work

• Mantpy — scverse framework for ECM analysis in spatial proteomics, which builds on thresholded ECM channels such as those produced here.

Contact

Mohamed Ghafoor — moeghaf@gmail.com