Lymphoma MVP (PFEE × HCL)

End-of-studies research project with the Hospices Civils de Lyon: predicting Hodgkin lymphoma treatment response from digitized histopathology slides.

Benchmarks four feature-extraction strategies on tissue patches: an ImageNet ResNet, a pathology foundation model, a from-scratch autoencoder, and I-JEPA — evaluated through linear probes with patient-level aggregation.

Explainability first: attribution maps trace every prediction back to the tissue regions that drove it — a hard requirement for clinical partners.

Built under clinical data-governance constraints (anonymization, auditability); 36 automated tests, TDD workflow. Code is private — no clinical data ever leaves the hospital pipeline.