From Images to Decisions: An 11-Month Journey of Transformation for a Chinese Medical AI Team

In late 2025, a paper from the preprint platform arXiv has been published and has attracted a lot of attention in the field of cardiovascular intervention community.

The paper reported that a domain-specific medical AI system named CA-GPT—designed to support decision-making during Percutaneous Coronary Intervention (PCI)—outperformed a general-purpose large language model (OpenAI's ChatGPT-5) across ten key clinical indicators. In some scenarios, its performance even exceeded that of junior physicians with one to five years of experience.

For the R&D team behind the system, however, the paper was not the destination.

What truly mattered was the long, uncertain, and sometimes uncomfortable path that led there.

This is the story of how a small medical AI team spent 11 months crossing the hardest gap in clinical AI: the distance between seeing information and making decisions.

From OCT Data to Clinical Decisions

Optical Coherence Tomography (OCT) is often called “the third eye” of interventional cardiologists. A thin optical fiber enters the coronary artery and produces high-resolution images that reveal plaque morphology and vessel structure in extraordinary detail.

For more than a decade, AI has been used to analyze OCT images in Vivolight OCT system, such as unstable plaque detection, calcified lesion assessment, and functional assessment. Many are already deployed in clinical workflows.

But something fundamental was missing. Doctors don’t just need more measurements. They need clear decisions.

Should a stent be implanted? Which diameter? What length? Is this lesion safe to treat, or should we step back?

Between raw imaging data and a real physician‘s decision lies a critical gap. No matter how accurate an algorithm is, if it cannot bridge that gap, it remains a tool—not a clinical partner.

That gap became the mission.

“The Last Mile” Is Always the Hardest

By early 2024, the Vivolight algorithm team had a clear vision: to turn years of expert knowledge into a system that could reason like an experienced physician—consistently, transparently, and under pressure.

They tried multiple approaches: general-purpose large models, API integrations, hybrid pipelines. Progress was slow. Each attempt felt close—but never quite enough.

As one team member put it:
——“You could feel you were approaching the answer. But every time, you were still one step away.”

A Turning Point: Stop Training Bigger Models

The breakthrough came in early 2025 with the open-sourcing of DeepSeek—not from training a larger model, but from asking a different question.

Instead of building another general AI, the team chose to anchor their system in what clinicians already trust: expert guidelines, industry consensus，structured reasoning, and explicit decision logic.

The idea was simple, but radical in practice:

● Let specialized vision models act as the “eyes”

● Let clinical guidelines and expert consensus form the “knowledge pool”

● Let a reasoning engine connect them into a transparent decision pathway

Learning the Hard Way—from Doctors

Early demos were humbling. Doctors didn’t ask how the AI worked. They didn’t care about prompts or architectures.

They asked questions like:

● “What stent would you choose?”

● “Is this lesion suitable for intervention?”

● “Why this diameter, not that one?”

When the system answered with vague explanations or generic language, the feedback was blunt. “This isn't AI. This is noise.” That comment from Dr. Zhu (CEO of Vivolight) hurt—but it clarified everything. Real clinical intelligence isn't about sounding smart. It’s about being precise, decisive, and accountable.

The team rebuilt again—this time modularly. Decision logic, reasoning steps, and outputs were stripped down, tested, and rebuilt at speed. Days blurred into nights. Iteration followed iteration.

Eventually, something changed.

Dr. Zhu looked at the system and said:
——“Okay. This actually feels like AI.”

Clinical Reality Is the Only Real Benchmark

By May 2025, the first version of AI-OCT system was delivered—less than three months after the project initiation. That same month, the innovation was showcased at the OCC 2025. To everyone's delight, it stood out among 48 entries and was selected as one of only five distinguished projects. But recognition was not the goal. Validation was.

The real test came from clinical trials. The system entered pilot use in multiple hospitals such as The Second Affiliated Hospital of Air Force Medical University (Tangdu Hospital), Beijing Anzhen Hospital Capital Medical University, Fuwai Hospital Chinese Academy of Medical Sciences, Zhongshan Hospital Affiliated to Fudan University, and The First Affiliated Hospital of Xi’an Jiaotong University, supporting PCI procedures in real-world settings. Feedback was direct, detailed, and often unforgiving.

Yet that feedback became the system’s strongest asset. Doctors challenged assumptions, corrected logic, and pushed the AI to be clearer, stricter, and more consistent. Step by step, the system matured—not in isolation, but in dialogue in clinical reality.