Artificial intelligence differentiates prefibrotic primary myelofibrosis with thrombocytosis from essential thrombocythemia using digitized bone marrow biopsy images

This repository contains the official code and model for the manuscript "Artificial intelligence differentiates prefibrotic primary myelofibrosis with thrombocytosis from essential thrombocythemia using digitized bone marrow biopsy images" (Srisuwananukorn et al., Leukemia 2026).

It provides a minimal, reproducible pipeline to:

1. Preprocess whole slide images (WSIs) into feature bags using the exact normalization parameters from the study.
2. Predict the probability of prePMF vs. ET using the trained Multiple Instance Learning (MIL) model.

Data Requirements

Annotation Manifest (Optional)

Provision of an annotation CSV is optional and only required if performance metrics (e.g., Accuracy, AUC) are desired. If provided, the file must contain the following columns:

Column	Description
slide	Required. The unique filename of the WSI (without extension).
bmbx_diagnosis	Required for Metrics. The ground truth label (prePMF or ET).

Example format:

slide,bmbx_diagnosis
slide_001,prePMF
slide_002,ET

Quick Start (Docker)

The provided Docker environment ensures a stable and reproducible execution of the pipeline.

1. Build the Docker Image

docker build -t prepmf_et_classifier .

2. Implementation: Preprocessing

The preprocess.py script automates tile extraction (302 µm) and feature generation (RetCCL) using study-specific normalization parameters.

docker run --gpus all \
    -v /path/to/wsis:/data \
    -v $(pwd)/output:/app/output \
    prepmf_et_classifier \
    python preprocess.py --wsi_dir /data --output_dir /app/output/features

3. Implementation: Prediction

The predict.py script executes the Attention-MIL model on the generated feature bags.

docker run --gpus all \
    -v $(pwd)/output:/app/output \
    prepmf_et_classifier \
    python predict.py --feature_dir /app/output/features --output_file /app/output/predictions.csv

Local Environment Setup

For execution in a local Python environment (3.8+), ensuring GPU support and appropriate drivers is recommended.

1. Dependency Installation:

   pip install -r requirements.txt

2. Preprocessing Workflow:

   python preprocess.py --wsi_dir /path/to/slides --output_dir ./features

3. Inference Workflow:

   python predict.py --feature_dir ./features --output_file predictions.csv

Model Architecture Details

• Feature Backbone: RetCCL (Residual Convolutional Curation and Learning) trained on large-scale histopathology datasets.
• Normalization: Macenko spectral normalization utilizing a fixed target stain matrix derived from the primary training cohort.
• Classifier: Attention-based Multiple Instance Learning (MIL) Aggregation.
• Resolution: 299 px tiles at 1.01 µm/px (~302 µm field-of-view).
• Environment: Optimized for Slideflow 2.2.1 (verified repo environment).

Citation

If you use this code or model in your research, please cite our original manuscript:

Srisuwananukorn A, Loscocco GG, Dolezal JM, Kuykendall AT, Santi R, Zhang L, Singh AM, Guglielmelli P, Vannucchi AM, Salama ME, Pearson AT, Hoffman R. Artificial intelligence differentiates prefibrotic primary myelofibrosis with thrombocytosis from essential thrombocythemia using digitized bone marrow biopsy images. Leukemia. 2026 Feb 20. doi: 10.1038/s41375-026-02893-7. Epub ahead of print. PMID: 41721088.