Alibaba's Damo Academy open sources RADAR, a medical vision-language model it says can read CT scans and identify ~150 abdominal conditions, including cancers

AI models can now read CT scans and identify nearly 150 abdominal conditions with accuracy exceeding many human radiologists.

Alibaba's Damo Academy has open-sourced RADAR, an AI model designed to read CT scans and identify approximately 150 abdominal conditions, including cancers. This vision-language model analyzes contrast-enhanced CT scans of 18 abdominal organs, correlating them with clinical reports to detect diseases and abnormalities. Trained on nearly 40,000 examinations, RADAR achieved an average AUC of 0.913 across 146 findings. In a comparative study published in Science, RADAR's average accuracy surpassed that of 23 out of 26 participating human radiologists. The model also assisted radiologists in improving missed diagnosis prevention by 10% and reducing analysis time by over 30%. Damo Academy suggests the training method could be extended to other medical imaging types, proposing RADAR as a generalist medical imaging model.

The open-sourcing of RADAR by Alibaba's Damo Academy signifies a growing trend of major tech companies contributing advanced AI tools to the medical field. This release positions RADAR as a potentially powerful, generalist medical imaging model, suggesting a future where AI assists or even surpasses human experts in diagnostic tasks across various imaging modalities. The model's strong performance, including outperforming a majority of human radiologists in a comparative study, highlights the rapid progress in AI's ability to interpret complex medical data. Companies are increasingly leveraging their AI expertise to address healthcare challenges, potentially accelerating the pace of innovation in medical diagnostics and patient care. This move by Alibaba could spur further research and development in this competitive space, encouraging other institutions to share their findings and models.

The success of RADAR, trained on a large dataset of CT scans and clinical reports, underscores the effectiveness of vision-language models in understanding nuanced medical information. Its ability to identify a wide array of conditions suggests a pathway towards more comprehensive AI diagnostic solutions that can handle multiple diseases simultaneously, rather than single-purpose tools. This has significant implications for healthcare accessibility and efficiency, particularly in regions with a shortage of specialized radiologists. The potential to extend this training methodology to other imaging types indicates a broader impact on medical AI development, promising more versatile AI assistants for clinicians. As these models become more capable and accessible, they could fundamentally alter the workflow and capabilities of diagnostic imaging departments worldwide.