Atlas 2 -- Foundation models for clinical deployment

2026-07-27 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Medical Devices & Health Technology, Health & Medical Research · Depth: Expert, extended

Summary

Atlas 2, Atlas 2-B, and Atlas 2-S are new pathology vision foundation models designed to overcome limitations in clinical deployment by achieving state-of-the-art performance, robustness, and resource efficiency. These models were trained on the largest pathology foundation model dataset to date, comprising 5.5 million histopathology whole slide images from Charité - Universitätsmedizin Berlin, LMU Munich, and Mayo Clinic. Atlas 2, a 2 billion parameter Vision Transformer, demonstrated superior performance, leading in 22 out of 27 tasks with an average 44.8% on HEST, 82.9% on eva, and 85.7% robustness across eighty public benchmarks. Its distilled versions, Atlas 2-B (86 million parameters) and Atlas 2-S (22 million parameters), are 24 and 91 times smaller, respectively, and 3.4 and 9 times more resource efficient. They also achieved top performance in their respective compute categories and showed significantly improved robustness.

Key takeaway

For computational pathology teams deploying AI models, you should consider the Atlas 2 family for clinical integration. These models demonstrate superior prediction performance, robustness to data variations, and resource efficiency across eighty benchmarks. Atlas 2-B and Atlas 2-S, in particular, offer highly efficient alternatives, being 24 and 91 times smaller in parameter size while maintaining top-tier performance. This minimizes tradeoffs, enabling more timely and cost-effective diagnostic analyses in routine clinical practice.

Key insights

Pathology foundation models can achieve state-of-the-art performance, robustness, and resource efficiency for clinical deployment through large-scale, multi-centric training and distillation.

Principles

• Multi-centric WSI datasets enhance pathology foundation model generalizability.
• Distillation yields efficient, high-performing pathology models for clinical use.
• Robustness to data variations is critical for clinical pathology AI.

Method

Train Vision Transformer models on 5.5 million histopathology WSIs at multiple resolutions (0.25, 0.5, 1.0, 2.0 microns/pixel). Distill larger models into smaller, efficient versions (e.g., ViT-B, ViT-S) for resource optimization.

In practice

• Consider Atlas 2-B or Atlas 2-S for efficient pathology AI inference.
• Evaluate models using comprehensive benchmarks like PathoROB for robustness.

Topics

• Pathology Foundation Models
• Clinical AI Deployment
• Whole Slide Imaging
• Vision Transformers
• Model Robustness
• Resource Efficiency

Code references

• kaiko-ai/eva
• mahmoodlab/HEST
• mahmoodlab/Patho-Bench
• bifold-pathomics/PathoROB
• owkin/plism-benchmark

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• A benchmark study of vision and pathology foundation models for computational pathology — Pathology-specific vision models perform strongly in computational pathology, but generalization varies, and ensembling improves aggregate performance.
• BoneCoT: multicentre validation of a whole-body skeleton foundation model for bone metastases guided by clinician-derived chain of thought — Clinician-derived reasoning can significantly enhance AI models for complex medical diagnostic tasks like bone metastasis detection.
• StrokeSeg2: Stroke Lesion Segmentation in Clinical Research Workflows — Architectural distillation significantly reduces deep learning model footprint and energy consumption for clinical deployment.
• Atlas H&E-TME: Scalable AI-Based Tissue Profiling at Expert Pathologist-Level Accuracy — Atlas H&E-TME provides scalable, expert-level AI-based tissue profiling from H&E slides, validated by a robust dual framework.
• DaX: Learning General Pathology Representations Across Scales — DaX is a pathology foundation model using DINOv3-style self-supervised learning to create robust, transferable visual representations across diverse scales and clinical tasks.
• Vision Foundation Models in Radiology: A Scoping Review of Data, Methodology, Evaluation and Clinical Translation — Radiology-specific Vision Foundation Models show promising transferability but face significant hurdles for clinical translation.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.