DaX: Learning General Pathology Representations Across Scales

Researchers present DaX, a pathology vision foundation model that adapts self-supervised learning to whole-slide histopathology imaging. The model demonstrates strong performance across a standardized benchmark of 161 clinical tasks, establishing a reproducible evaluation framework for computational pathology applications.

DaX represents a significant advancement in computational pathology by addressing a critical challenge: creating visual representations that maintain robustness across diverse clinical environments. Traditional pathology image analysis struggles with variability in magnification levels, staining protocols, scanner manufacturers, and slide preparation methods. The DaX model tackles this through multiple architectural innovations, including continuous magnification training, cross-scale tissue views, and orientation-agnostic augmentation techniques. This enables the system to connect cellular-level morphology with broader tissue architecture while maintaining consistent performance across input variations.

The research builds on momentum in medical AI, where foundation models trained on diverse datasets increasingly outperform task-specific approaches. DaX's development from natural-image DINOv3 weights demonstrates effective transfer learning from general computer vision to specialized medical domains. The model's Gram-anchored dense consistency mechanism specifically addresses the challenge of stabilizing token-level representations across different scales, a problem unique to histopathology where analysis spans multiple magnification levels.

The construction of a comprehensive WSI-level benchmark spanning 44 public datasets and over 28,000 patients provides the field with standardized evaluation infrastructure. This addresses a persistent issue in medical AI: published results often lack consistency because different teams use different validation protocols. The patient-level cross-validation with fold-level statistical ranking reduces noise from dataset-dependent variation, enabling more reliable model comparisons.

For clinical adoption, DaX's consistent performance across diagnostic pathology, biomarker profiling, and prognostic tasks suggests broad utility in real-world settings. The open benchmark framework may accelerate development of subsequent pathology models by providing a common evaluation standard, similar to how ImageNet transformed general computer vision research.

• →DaX achieves highest mean performance across 161 clinical tasks from 44 public datasets covering 28,182 patients
• →The model incorporates continuous magnification training and cross-scale consistency to handle pathology imaging variability
• →Standardized WSI-level benchmark with patient-level cross-validation enables reproducible comparisons across pathology foundation models
• →Strong performance spans diagnostic pathology, biomarker profiling, tissue context, and prognostic prediction tasks
• →Transfer learning from natural-image DINOv3 weights demonstrates effective domain adaptation to histopathology