A fully open AI foundation model applied to chest radiography

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Radiology
• Foundation Models in Healthcare

Background:

• Chest radiography is a primary tool for diagnosing lung diseases.
• Current deep learning models for chest X-rays have limitations in scope, generalizability, and robustness.
• Need for advanced AI solutions to improve diagnostic accuracy and efficiency in radiology.

Purpose of the Study:

• To develop Ark+, a foundation model for chest radiography to overcome limitations of existing deep learning models.
• To enhance diagnostic scope, generalizability, adaptability, and robustness in thoracic disease interpretation.
• To create an open-source, extensible AI model for medical imaging.

Main Methods:

• Developed Ark+ using a foundation model approach for chest radiography.
• Pretrained the model by cyclically accruing and reusing knowledge from heterogeneous expert labels across numerous datasets.
• Incorporated strategies to handle data biases, long-tailed distributions, and support federated learning for privacy.

Main Results:

• Ark+ demonstrates superior performance in diagnosing thoracic diseases, expanding diagnostic scope and reducing misdiagnosis.
• The model shows adaptability to evolving diagnostic needs, can learn rare conditions from few samples, and transfers knowledge without retraining.
• Achieved unprecedented performance by aggregating diverse datasets and expert knowledge, surpassing proprietary models.

Conclusions:

• Ark+ represents a significant advancement in AI for chest radiography, offering improved diagnostic capabilities and adaptability.
• Open foundation models trained on diverse, heterogeneous data can outperform proprietary models, promoting open science in AI for medicine.
• The release of Ark+ code and models facilitates fine-tuning, local adaptation, and further improvement, democratizing AI in healthcare.