Arizona State University researchers have developed Ark+, an open-source artificial intelligence tool for chest X-ray diagnosis that outperforms proprietary models from Google and Microsoft. The system demonstrates superior accuracy across multiple thoracic conditions while offering full accessibility to healthcare providers worldwide.
In a David versus Goliath scenario that has reshaped medical AI development, a small research team at Arizona State University has created an artificial intelligence system that significantly outperforms the proprietary chest X-ray analysis tools developed by technology giants Google and Microsoft.
The breakthrough model, called Ark+, represents a paradigm shift in medical imaging AI by proving that open-source development using publicly available datasets can surpass the performance of closed, proprietary systems trained on massive private datasets.
“Ark+ is designed to be an open, reliable and ultimately useful tool in real-world health care systems,” said Professor Jianming “Jimmy” Liang from ASU’s College of Health Solutions, lead author of the study published in Nature on 11 June 2025.
Superior diagnostic capabilities demonstrated
The research team evaluated Ark+ across eight clinical scenarios using ten different datasets, demonstrating exceptional performance in diagnosing both common and rare thoracic diseases. In internal evaluations using the established ChestXray14 dataset, Ark+ achieved a mean area under the curve (AUC) of 84.43%, significantly outperforming competing models including RAD-DINO (83.54%), MIM-CXR (83.08%), and CheSS (80.46%).
Perhaps most remarkably, Ark+ demonstrated superior zero-shot transfer capabilities, achieving impressive AUC scores for detecting pneumothorax (95.79%), paediatric pneumonia (97.60%), and tuberculosis (96.60%) in datasets it had never encountered during training.
The system’s diagnostic scope extends beyond traditional AI limitations. Through its multi-task head architecture, Ark+ can identify conditions across different classification systems simultaneously, expanding diagnostic capabilities and addressing potential misdiagnoses by human experts.
Innovative training methodology drives performance
The key to Ark+’s superior performance lies in its novel training approach, which the researchers term “cyclically accruing and reusing knowledge.” Unlike conventional AI systems that typically use self-supervised learning methods, Ark+ was trained on heterogeneous expert labels from six public datasets comprising over 700,000 worldwide images.
“We wanted AI to learn from expert knowledge, not only from the raw data,” explained Liang. This approach preserves and utilises the most valuable information in medical datasets – the detailed diagnostic annotations provided by radiologists and other medical experts.
The system employs a teacher-student framework with multi-task heads, where knowledge is continuously accumulated and reused across different diagnostic tasks. This cyclical pretraining process enables the model to leverage expertise from multiple institutions and diagnostic approaches.
Addressing critical healthcare challenges
Ark+ tackles several fundamental challenges in medical AI deployment. The system demonstrates remarkable few-shot learning capabilities, successfully diagnosing rare conditions like subcutaneous emphysema, tortuous aorta, and pneumoperitoneum using only 1-5 training samples per condition.
The model also exhibits robust performance in handling long-tailed disease distributions, where some conditions are significantly more common than others. In evaluations using the ChestDR dataset with 19 thoracic diseases exhibiting long-tailed distributions, Ark+ achieved a mean AUC of 86.55%, substantially outperforming other foundation models.
Crucially, Ark+ demonstrates resilience to sex-related bias, showing superior performance across demographic groups without requiring specific bias mitigation techniques. In comprehensive bias testing, Ark+ produced 13 unbiased results compared to just 4-8 for competing models.
Adaptability to emerging health threats
The system’s extensibility was demonstrated through its response to novel diseases. Despite having no prior exposure to COVID-19 during training, Ark+ achieved competitive diagnostic accuracy for COVID-19 detection. Through incremental learning, the model was successfully updated to create Ark++covid, which surpassed existing models in COVID-19 diagnosis.
“By making this model fully open, we’re inviting others to join us in making medical AI more fair, accurate and accessible,” Liang added. This adaptability positions Ark+ as a valuable tool for responding to future pandemic scenarios.
Open-source philosophy drives global impact
Unlike proprietary alternatives, Ark+ is fully open-source, with all code and pretrained models publicly available. This approach enables researchers worldwide to fine-tune, adapt, and improve the system for local clinical needs.
The model can be extended through federated learning to incorporate private data while preserving patient privacy, addressing a critical concern in medical AI deployment. The researchers demonstrated that federated Ark+ maintains high performance while enabling collaborative development across multiple institutions without compromising data security.
Transforming medical imaging accessibility
Ark+’s open-source nature addresses a critical equity issue in healthcare AI. While proprietary models restrict access and limit customisation, Ark+ can be deployed in resource-limited settings and adapted for specific population needs.
The system’s relatively modest computational requirements – training takes approximately 700 hours using four A100 GPUs – make it accessible to smaller institutions and developing healthcare systems. This democratisation of advanced diagnostic AI could significantly impact global health outcomes.
Clinical validation and real-world applications
The research included comprehensive clinical validation through collaboration with Mayo Clinic Arizona. Senior radiologists reviewed cases where Ark+ disagreed with official dataset labels, confirming the system’s ability to identify previously missed conditions and correct potential overdiagnoses.
The model’s performance extended beyond classification to support device identification, with successful detection of support devices like PICC lines and internal jugular catheters, expanding its clinical utility.
Future implications for medical AI
The success of Ark+ challenges the assumption that proprietary models with access to vast private datasets necessarily outperform open-source alternatives. The research demonstrates that strategic use of public datasets combined with innovative training methodologies can achieve superior results.
As the authors note in their discussion: “The development of Ark+ reveals that open models trained by accruing and reusing knowledge from heterogeneous expert annotations with a multitude of public (big or small) datasets can surpass the performance of proprietary models trained on large data.”
This paradigm shift could accelerate medical AI development by encouraging greater data sharing and collaborative model development, ultimately benefiting global healthcare delivery.
The research represents a significant milestone in opening access to advanced medical AI, offering healthcare providers worldwide access to state-of-the-art diagnostic capabilities while maintaining the flexibility to adapt systems to local needs and emerging health challenges.
Reference:
Ma, D., Pang, J., Gotway, M. B., & Liang, J. (2025). A fully open AI foundation model applied to chest radiography. Nature, 643, 488-498.
https://doi.org/10.1038/s41586-025-09079-8
