Vision Foundry: A System for Training Foundational Vision AI Models.

Area of Science:

• Medical AI
• Computer Vision
• Machine Learning

Background:

• Self-supervised learning (SSL) offers potential for medical imaging analysis but faces adoption barriers due to technical complexity.
• Clinical researchers require accessible tools to leverage large, unannotated medical datasets for AI model development.

Purpose of the Study:

• To introduce Vision Foundry, a code-free, HIPAA-compliant platform simplifying the pre-training, adaptation, and deployment of foundational vision models for clinical applications.
• To enable domain experts to develop advanced clinical AI tools with reduced annotation needs.

Main Methods:

• The Vision Foundry platform integrates the DINO-MX framework, handling distributed infrastructure complexities.
• It implements specialized strategies including Magnification-Aware Distillation (MAD) and Parameter-Efficient Fine-Tuning (PEFT).
• Validation was performed across neuropathology segmentation, lung cellularity estimation, and coronary calcium scoring tasks.

Main Results:

• Models trained using Vision Foundry demonstrated superior segmentation fidelity and regression accuracy compared to generic baselines.
• The platform facilitated robust zero-shot generalization across diverse imaging protocols.
• Vision Foundry significantly reduced the annotation overhead for developing state-of-the-art clinical AI.

Conclusions:

• Vision Foundry democratizes advanced representation learning for clinical researchers, lowering technical barriers to SSL adoption.
• The platform empowers domain experts to focus on clinical discovery rather than engineering optimization.
• It facilitates the development of high-performing, generalizable clinical AI tools from unannotated medical data.