MADELEINE

MADELEINE is a slide-level foundation model for computational pathology that learns whole-slide image (WSI) representations through multistain pretraining. Most self-supervised approaches to slide representation learning construct positive pairs from augmented views of a single slide. MADELEINE instead treats different histochemical and immunohistochemical stains of the same tissue as naturally occurring views: hematoxylin and eosin (H&E) alongside marker stains that highlight complementary morphological and molecular features. Aligning these stains during pretraining forces the model to capture tissue biology that is invariant to staining protocol, yielding slide embeddings that transfer well across downstream tasks.

The model was developed by the Mahmood Lab at Harvard Medical School and Brigham and Women's Hospital and introduced by Guillaume Jaume, Anurag Vaidya, Andrew Zhang, Andrew H. Song, Richard J. Chen, Sharifa Sahai, Dandan Mo, Emilio Madrigal, Long Phi Le, and Faisal Mahmood at ECCV 2024 (arXiv:2408.02859). It builds directly on the lab's patch-level encoders: MADELEINE consumes patch features from CONCH and aggregates them into a single slide embedding, extending the Mahmood Lab pathology stack from the tile level (UNI, CONCH) up to the whole-slide level.

MADELEINE targets a practical bottleneck in digital pathology, where a gigapixel slide can contain tens of thousands of patches and slide-level labels are scarce. By pretraining on paired stains and evaluating on H&E at inference, it produces compact slide representations that support few-shot and low-data downstream modeling.

#Key Features

• Multistain pretraining: Uses multiple stains of the same tissue section as paired views, learning representations that are robust to staining protocol rather than relying on synthetic augmentations of a single slide.
• Dual global-local alignment: Combines a global InfoNCE contrastive objective that aligns whole-slide embeddings across stains with a local Graph Optimal Transport objective that matches patch-level distributions and graph structure between stains.
• Attention-based slide aggregation: A multihead attention-based multiple instance learning (ABMIL) network aggregates thousands of patch features into a single fixed-length slide embedding.
• Built on CONCH features: Patches are encoded with the frozen CONCH image encoder at 256x256 pixels, so MADELEINE learns aggregation rather than re-learning patch-level morphology from scratch.
• H&E-only inference: Although pretraining requires paired stains, the deployed model produces slide embeddings from H&E slides alone, matching how most clinical archives are stained.

#Technical Details

MADELEINE was pretrained on two clinical cohorts: 4,211 breast cancer WSIs from 1,153 cases across five stains (H&E, ER, PR, HER2, KI67, processed at 10x) and 12,070 kidney transplant WSIs from 1,069 cases across four stains (H&E, PAS, Jones, Trichrome, processed at 20x), totaling 16,281 slides. Each slide is tiled into 256x256-pixel patches encoded by CONCH; an ABMIL aggregator with multiple attention heads produces the slide embedding. The global objective aligns embeddings of matched stains via symmetric InfoNCE, while the local objective uses Graph Optimal Transport, minimizing the Wasserstein distance between patch embedding distributions and the Gromov-Wasserstein distance between their graph structures. The learned representations were evaluated on 21 downstream tasks using 7,299 WSIs from multiple medical centers, spanning morphological classification, molecular status prediction, and prognostic prediction. In few-shot molecular status prediction, MADELEINE reached 0.838 AUC for estrogen receptor status at k=25 labeled slides per class, compared with 0.762 for GigaPath.

#Applications

MADELEINE serves as a slide-level feature extractor for computational pathology pipelines where slide-level labels are limited. Researchers can embed H&E whole slides and train lightweight classifiers for cancer subtyping, biomarker and molecular status prediction (such as ER, PR, and HER2 in breast cancer), and survival or prognostic modeling, often with only a handful of labeled slides per class. Because pretraining explicitly aligns marker stains with H&E morphology, the resulting embeddings are particularly suited to tasks that connect routine H&E appearance to molecular readouts otherwise assessed with specialized stains.

#Impact

MADELEINE demonstrates that the multiple stains routinely produced in clinical workflows are a rich, label-free supervisory signal for slide representation learning, an alternative to augmentation-based self-supervision on single slides. It complements the Mahmood Lab's tile-level encoders by providing a slide-level aggregator, and its ABMIL-plus-optimal-transport recipe informs subsequent slide foundation models. Adoption is shaped by its release terms: the pretrained weights are distributed on HuggingFace under an MIT license but gated behind a contact-sharing agreement, while the training and evaluation code is released under the non-commercial, no-derivatives CC BY-NC-ND 4.0 license, and the clinical pretraining cohorts are not public. The model has not been validated for clinical diagnostic use and requires independent evaluation before any diagnostic application.