Foundation Ark: Accruing and Reusing Knowledge for Superior and Robust Performance.

Area of Science:

• Medical Imaging
• Deep Learning
• Artificial Intelligence

Background:

• Deep learning models require massive annotated datasets for expert-level performance.
• Existing public medical imaging datasets are often small and heterogeneous in expert labels.
• Proprietary models, like Google's CXR-FM, are trained on large, private datasets.

Purpose of the Study:

• To develop a framework (Ark) for training powerful foundation models by aggregating numerous small, public datasets.
• To leverage heterogeneous expert annotations from diverse sources.
• To demonstrate that aggregated public data can outperform large, private datasets.

Main Methods:

• Developed the Ark framework to accrue and reuse knowledge from heterogeneous expert annotations.
• Trained two Ark models by merging multiple public chest X-ray datasets (ChestX-ray14, CheXpert, MIMIC-II, VinDr-CXR).
• Evaluated models on classification and segmentation tasks using fine-tuning, linear-probing, and gender-bias analysis.

Main Results:

• Ark models demonstrated superior and robust performance over SOTA baselines and Google's CXR-FM.
• Aggregating public datasets diversified patient populations and expert knowledge, enhancing performance.
• Achieved high performance while saving on annotation costs.

Conclusions:

• Ark framework effectively utilizes aggregated public datasets for training high-performing medical deep learning models.
• Open science initiatives, sharing data and code, can lead to foundation models surpassing proprietary ones.
• Ark promotes the democratization of deep learning in medical imaging through open access to code and models.