Ark⁺: Supervised training a single high-performance AI foundation model from many differently labeled datasets-no...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Deep learning models require massive labeled datasets for high performance, often necessitating proprietary data.
• Publicly available medical imaging datasets are numerous but individually small and suffer from heterogeneous expert labels.
• Existing methods struggle to aggregate these diverse datasets for training a single, robust AI model.

Purpose of the Study:

• To introduce Ark+, a novel framework for training a single, high-performance artificial intelligence (AI) model using multiple, heterogeneously labeled datasets.
• To overcome the challenge of label heterogeneity in supervised learning without manual data harmonization.
• To demonstrate the capability of Ark+ in creating robust AI foundation models applicable across various domains.

Main Methods:

• Developed Ark+, a framework designed to accrue and reuse knowledge from heterogeneous expert annotations across diverse datasets.
• Pretrained Ark+ models (Ark+5 and Ark+6) on aggregated public chest radiograph datasets (e.g., MIMIC-CXR, CheXpert).
• Evaluated Ark+ models on classification, segmentation, and localization tasks, including ablation studies and simulations for federated learning and multi-modality applications.

Main Results:

• Ark+ models demonstrated superior and robust performance compared to state-of-the-art baselines and proprietary models (Google's CXR-FM).
• Ablation studies confirmed the effectiveness of Ark+'s components and its advantage over alternative strategies.
• Ark+ showed scalability, independence of architecture, and extensibility to different imaging modalities (e.g., fundus photography).

Conclusions:

• Ark+ offers a methodological breakthrough in supervised learning, enabling the creation of powerful AI foundation models from diverse, publicly available data.
• The framework effectively handles label heterogeneity, reduces annotation costs, and diversifies patient populations for enhanced model performance.
• Ark+ has significant implications for open science, promoting the development of open, superior, and robust foundation models across various scientific disciplines.