Atlas 2 - Foundation models for clinical deployment

Maximilian Alber, Timo Milbich, Alexandra Carpen-Amarie, Stephan Tietz, Jonas Dippel, Lukas Muttenthaler, Beatriz Perez Cancer, Alessandro Benetti, Panos Korfiatis, Elias Eulig, Jérôme Lüscher, Jiasen Wu, Sayed Abid Hashimi, Gabriel Dernbach, Simon Schallenberg, Neelay Shah, Moritz Krügener, Aniruddh Jammoria, Jake Matras, Patrick Duffy, Matt Redlon, Philipp Jurmeister, David Horst, Lukas Ruff, Klaus-Robert Müller, Frederick Klauschen, Andrew P. Norgan

TL;DR

The authors introduce Atlas 2, a state-of-the-art pathology vision foundation model trained on a massive multi-centric dataset of 5.5 million whole slide images. Alongside the main 2-billion-parameter model, they release distilled, lightweight variants (Atlas 2-B and Atlas 2-S) that break traditional performance-efficiency trade-offs, delivering clinical-grade accuracy and unprecedented robustness at up to 9x the processing speed.

What Problem Does This Paper Address?

While AI has shown immense promise in digital pathology, moving models from research to actual clinical deployment has faced a triple bottleneck:

  1. Robustness and Generalization: Histopathology images vary wildly depending on the lab, the staining protocol, and the scanner hardware used. Models trained in one hospital often fail when deployed in another.
  2. Data Scarcity: Gathering enough labeled data for rare diseases is logistically and financially prohibitive.
  3. Computational Efficiency: Whole slide images (WSIs) are gigapixel-scale files. Processing them with massive frontier models requires significant compute, making real-time clinical workflows expensive and slow.

The authors address these challenges by scaling up both dataset diversity and model capacity, while simultaneously using knowledge distillation to create highly efficient models tailored for real-world clinical infrastructure.

Methods

The core architecture relies on the Vision Transformer (ViT) with a fine-grained patch-token size of 8. The training pipeline progresses through three tiers:

  • Atlas 2 (The Frontier Model): A massive 2-billion-parameter ViT variant trained using an optimized self-supervised learning recipe drawing from RudolfV, DINOv2, and DINOv3 frameworks. It processes image tiles across multiple magnifications (5×, 10×, 20×, and 40×) to capture both broad tissue architecture and fine cellular details.
  • Atlas 2-B & Atlas 2-S (The Clinical Workhorses): To make these capabilities practical, the authors distilled the knowledge from the 2B model into smaller, resource-efficient architectures: ViT-Base (86 million parameters) and ViT-Small (22 million parameters).

Datasets

Training Data

The models were trained on the largest multi-centric pathology dataset to date: 5.5 million de-identified histopathology whole slide images. This corpus was aggregated from three premier medical institutions to ensure technical and biological diversity:

  • Charité - Universitätsmedizin Berlin
  • LMU Munich
  • Mayo Clinic

Evaluation Benchmarks

Rather than testing on a handful of tasks, the authors evaluated their models across eighty public benchmark datasets consolidated under five rigorous evaluation frameworks:

  • HEST: For spatial transcriptomics and gene expression prediction.
  • eva: For clinical prediction and patch/slide-level classification.
  • PathoROB: Specifically designed to measure model robustness against non-biological confounders (like hospital-specific artifacts).
  • Plismbench: To test representation consistency across different scanners and stains.
  • Patho-Bench: Covering 53 tasks across morphology, molecular subtyping, treatment response, and survival prediction.

Key Insights and Ideas

  • Multi-Centric Data Inherent Resolves Batch Effects: By training on millions of slides from different hospital networks across Europe and the US, the model naturally learns to ignore confounding factors like variations in scanner hardware or staining chemicals.
  • The Power of Distillation in Pathology: Bigger is better for learning representations, but smaller is better for deployment. The paper proves that knowledge distillation allows a 22M or 86M parameter model to retain the edge-case accuracy of a 2B parameter model.
  • Redefining the Pareto Front: Atlas 2-B and Atlas 2-S effectively establish a new state-of-the-art Pareto front for performance vs. speed and performance vs. robustness, outperforming existing models that are orders of magnitude larger.

Key Results

The comprehensive evaluation solidifies Atlas 2 as the current leading pathology foundation model:

  • Top-Tier Performance: Atlas 2 achieved the best performance in 22 out of 27 primary benchmark tasks, outperforming other prominent models like Pluto-4G and UNI2-H.
  • Unmatched Robustness: On the PathoROB Robustness Index, Atlas 2 scored 85.7%, marking a massive 9.7 percentage point improvement over the closest competitor (Virchow2).
  • Dramatic Speedups: The distilled models, Atlas 2-B and Atlas 2-S, process standard-sized images 3.4× and 9× faster than the main model, respectively. Remarkably, Atlas 2-S matches or exceeds the performance of much larger models (like Virchow2 and Midnight-12k) while operating at a fraction of the computational cost.
Figure 1: (A) The results show that Atlas 2 is the best performing model. Additionally Atlas 2-B and 2-S have similar prediction performance as contenders but are up to a magnitude more efficient. The performance average is based on morphology prediction tasks from eva [27] and molecular prediction tasks from HEST [34] in Table 1. The dots show the processing speed of the respective model for a 224×224 pixel image on an L4 GPU. The value for Pluto-4G [60] is approximated by taking the fastest speed of same sized models [38, 67, 79]. (B) The results show that Atlas 2, Atlas 2-B, and Atlas 2-S are the most robust models and also the state of the art performance-robustness Pareto front. The plot shows the performance average from (A) and the robustness average from Table 1. (C) The data shows that the Atlas 2 models are the performance-resource efficiency Pareto front. The plot shows the performance average of (A).

Limitations and Future Work

Despite the sweeping success across classification and robustness benchmarks, the authors highlight a persistent wall in survival prediction. Across all evaluated models (including Atlas 2), performance on survival tasks remained close to chance level (C-Index around 50). This indicates that predicting long-term patient survival directly from localized tissue morphology remains an incredibly complex bottleneck that scaling up visual data alone cannot fully solve. Future work will likely need to focus on better multi-modal integration (combining clinical text and genomics) or advanced slide-level aggregation techniques.

Summary updated

Abstract

Pathology foundation models substantially advanced the possibilities in computational pathology -- yet tradeoffs in terms of performance, robustness, and computational requirements remained, which limited their clinical deployment. In this report, we present Atlas 2, Atlas 2-B, and Atlas 2-S, three pathology vision foundation models which bridge these shortcomings by showing state-of-the-art performance in prediction performance, robustness, and resource efficiency in a comprehensive evaluation across eighty public benchmarks. Our models were trained on the largest pathology foundation model dataset to date comprising 5.5 million histopathology whole slide images, collected from three medical institutions Charité - Universtätsmedizin Berlin, LMU Munich, and Mayo Clinic.
BibTeX
@misc{alber2026atlas,
  author = {Maximilian Alber and Timo Milbich and Alexandra Carpen-Amarie and Stephan Tietz and Jonas Dippel and Lukas Muttenthaler and Beatriz Perez Cancer and Alessandro Benetti and Panos Korfiatis and Elias Eulig and Jérôme Lüscher and Jiasen Wu and Sayed Abid Hashimi and Gabriel Dernbach and Simon Schallenberg and Neelay Shah and Moritz Krügener and Aniruddh Jammoria and Jake Matras and Patrick Duffy and Matt Redlon and Philipp Jurmeister and David Horst and Lukas Ruff and Klaus-Robert Müller and Frederick Klauschen and Andrew P. Norgan},
  title = {{Atlas 2 - Foundation models for clinical deployment}},
  year = {2026},
  eprint = {2601.05148},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.05148},
}
Figure