RudolfV

RudolfV is a self-supervised foundation model for computational pathology developed by Aignostics together with academic collaborators including TU Berlin, BIFOLD, Charité – Universitätsmedizin Berlin, the German Cancer Research Center (DKFZ), and LMU Munich. Introduced in January 2024, the model is named after Rudolf Virchow, a founder of modern pathology, reflecting its design philosophy: a foundation model built "by pathologists for pathologists," in which domain expertise guided data curation and evaluation rather than relying purely on scale.

Histopathology slides are extremely heterogeneous, spanning many tissue types, disease entities, staining protocols, and scanner vendors, and most computational pathology models struggle to generalize across this variation or to handle rare diseases. RudolfV addresses this by combining a large, deliberately diverse training corpus with self-supervised pretraining, producing a general-purpose tile encoder whose embeddings transfer to a wide range of downstream diagnostic and biomarker tasks.

The model sits alongside other pathology foundation models such as UNI, Virchow, H-optimus-0, and Hibou, and was among the early efforts to emphasize stain and laboratory diversity as a deliberate design axis rather than simply maximizing the number of hematoxylin-and-eosin slides.

#Key Features

• Pathologist-guided curation: Data selection and the evaluation framework were shaped by practicing pathologists, prioritizing tissue, disease, and staining diversity over raw slide count.
• Multi-stain coverage: Training data spans roughly 97 unique staining types, including H&E (about 70%), immunohistochemistry (about 10%), and other histochemical stains, broadening applicability beyond H&E-only models.
• Cross-laboratory diversity: Slides were sourced from more than 15 laboratories across the EU and US, covering 58 tissue types to improve robustness to scanner and protocol variation.
• Strong downstream transfer: As a frozen feature extractor, RudolfV matches or surpasses contemporary foundation models on tile-level benchmarks and tumor microenvironment and biomarker tasks.

#Technical Details

RudolfV is a Vision Transformer (ViT-L/14, roughly 300 million parameters) pretrained with a DINOv2-style self-supervised objective. The training adaptation samples a specific distribution over slide groups and tissue clusters and extends the standard augmentation pipeline with stain variations to encourage stain-invariant representations. The corpus comprises 103,849 whole-slide images from 35,784 cases, from which about 791 million tiles were extracted and roughly 751 million retained after filtering. Pretraining used a batch size of 960 on 16 A100-40GB GPUs for 625,000 iterations. Evaluated as a frozen encoder across benchmarks including PCam, MHIST, CRC-100K, MSI prediction in colorectal and gastric cancer, and tumor-infiltrating-lymphocyte detection, RudolfV reports competitive or state-of-the-art performance relative to other foundation models of its era while using comparatively fewer slides.

#Applications

RudolfV serves as a backbone for computational pathology workflows in both clinical research and biopharma. Its embeddings support tasks such as cancer subtyping, nuclear and tissue segmentation, microsatellite-instability and other biomarker prediction, and tumor microenvironment profiling, typically by training lightweight heads on the frozen features. Aignostics has described RudolfV as the base model underlying its histopathology product work, making it relevant to diagnostic-support tooling, translational research, and clinical-trial biomarker analysis.

#Impact

RudolfV helped establish data diversity and pathologist-informed curation, rather than slide count alone, as decisive factors for pathology foundation models, demonstrating competitive performance from a curated multi-stain corpus. It serves as the predecessor to Aignostics' later Atlas model (developed with Mayo Clinic and Charité) and is frequently cited in surveys of computational pathology foundation models. Practical adoption outside Aignostics is constrained by access terms: the work is released under a CC BY-NC-ND 4.0 license, and weights are distributed through Aignostics rather than as a fully open release, which limits broad academic reuse compared with openly licensed alternatives.