Voio, a frontier AI lab dedicated to healthcare, has emerged from stealth with $8.6 million in seed funding.
The company, founded by leading researchers from the University of California, Berkeley (UC Berkeley) and the University of California, San Francisco (UCSF), is building a unified platform to assist radiologists. Simultaneously, the team has launched Pillar-0, an open-source artificial intelligence (AI) model, which the founding researchers boast is a highly accurate model for medical imaging, outperforming models from Google, Microsoft, and Alibaba.
The $8.6 million seed round was co-led by Laude Ventures, co-founded by the founder of Databricks, and Perplexity, and the UC Berkley’s venture capital firm, The House Fund. The funding will support Voio’s mission to scale its impact across radiology, developing richer AI systems that can collaborate across medical specialties and push the frontier of predictive, preventative care.
Voio’s mission is to create a unified reading environment that integrates image viewing, reporting, and AI tools into a single, intelligent platform.
“Radiologists shouldn’t have to choose between speed and quality,” said Dr. Maggie Chung, Co-Founder and Medical Lead of Voio. “Our goal is to make radiology reporting seamless by drafting full reports and connecting images, history, and prior exams into one intelligent platform.”
Outperforming Industry Giants
The research team validated Pillar-0 on a diverse set of scans, including chest CT, abdomen CT, brain CT, and breast MRI. The model achieved an average Area Under the Curve (AUC) score of .87 across more than 350 different clinical findings. This metric, where 1.0 represents a perfect test, demonstrates a substantial 10% to 17% improvement in accuracy over leading publicly available models from Google (MedGemma, .76 AUC), Microsoft (MI2, .75 AUC), and Alibaba (Lingshu, .70 AUC).
“Pillar-0 marks a major milestone in our mission to push the frontier of AI for radiology,” said Adam Yala, Assistant Professor at UC Berkeley and UCSF and a senior author of the research. He noted that the model not only outperforms competitors but also runs an order of magnitude faster and can be fine-tuned with minimal effort for specific clinical tasks.
How It Works
A key innovation of Pillar-0 is its ability to process entire 3D imaging volumes directly, unlike many existing models that analyze 2D slices independently. This is enabled by a novel neural network architecture called “Atlas,” which the developers say is over 150 times faster than traditional vision transformers when processing an abdomen CT scan.
This efficiency allows Pillar-0 to serve as a powerful, general-purpose backbone. When fine-tuned for specific applications, it has already shown remarkable results. In an external validation study at Massachusetts General Hospital, a version of Pillar-0 improved upon the state-of-the-art lung cancer prediction tool, Sybil-1, by 7%. For brain hemorrhage detection, it outperformed all baseline models while using only a quarter of the training data.
Emphasizing a commitment to open science, the team has released the complete Pillar-0 codebase, trained models, and evaluation pipelines to the public. They have also introduced a new clinically-grounded evaluation framework called RaTE, designed to provide a more realistic measure of model utility compared to existing benchmarks.
“Transparency is essential to advancing the science of AI in health,” said Yala. “Open-sourcing enables the global research community to independently validate our tools and build on our work.” The team plans to expand the model’s capabilities to additional imaging modalities and eventually to full report generation.
