aliceblondel/mibc-ai-subtyping

MIBC AI Subtyping

AI models for molecular subtyping of muscle-invasive bladder cancer (MIBC) from H&E whole-slide images.

Paper: Deep Learning Bridges Histology and Transcriptomics to Predict Molecular Subtypes and Outcomes in Muscle-Invasive Bladder Cancer

What this model does

Given H&E WSI tiles encoded with H-optimus-1, the model predicts:

• Consensus molecular subtype: Ba.Sq, LumU, LumP, LumNS, Stroma.rich
• Bulk gene expression (log2 RPM+1) from histology
• Optionally detects NMIBC / Non-Tumor slides before subtyping (use_mibc_detect)

Model weights

This repository contains:

• MIBCSubtyping_checkpoints/ — 10-fold ensemble of BulkMIL models (subtype + gene expression prediction)
• MIBCDetect_checkpoints/ — 10-fold ensemble of tile classifiers (MIBC / NMIBC / Non-Tumor detection)

Citation

@article{Blondel2025.10.23.684013,
  author = {Blondel, Alice and Krucker, Clémentine and Harter, Valentin and Da Silva, Melissa and Groeneveld, Clarice S. and de Reynies, Aurélien and Karimi, Maryam and Benhamou, Simone and Bernard-Pierrot, Isabelle and Pfister, Christian and Culine, Stéphane and Allory, Yves and Walter, Thomas and Fontugne, Jacqueline},
  title = {Deep Learning Bridges Histology and Transcriptomics to Predict Molecular Subtypes and Outcomes in Muscle-Invasive Bladder Cancer},
  journal = {bioRxiv},
  year = {2025},
  doi = {10.1101/2025.10.23.684013},
}