Shanghai Pulmonary Hospital advances pathology Vibe Research with PathEngine.

From Vibe Coding to Vibe Research in pathology

PathEngine brings Vibe Research into pathology: pathologists express clinical intent in natural language, define the disease problem and use the platform to build models without writing code.

With PathEngine, a clinical or research idea can become a model-building task. The physician provides the medical question and domain judgment, while the platform helps organize data, train, evaluate and deploy the model.

A pathology foundation model as the base

PathOrchestra was released with leading pathology experts as China's first pathology foundation model. It was trained on 287,424 whole-slide images across 21 organ categories from authoritative centers, including Xijing Hospital, the First Affiliated Hospital of USTC and TCGA.

The model was evaluated on 112 clinically relevant tasks, 27,755 WSI and 9.41 million ROI images. Across 47 tasks, accuracy exceeded 0.95, making the foundation model a reusable base for downstream pathology AI development.

PathEngine as a zero-code AI production line

Pathology AI faces a long-standing contradiction: clinical demand is industrial in scale, but model development often remains workshop-like. Traditional pathology AI development can take 12-24 months and requires large labeled datasets and algorithm engineers.

PathEngine addresses this by connecting data management, annotation, training, evaluation and release in a zero-code workflow. It integrates PathOrchestra and DaYi so hospitals can build specialty models for long-tail pathology questions more efficiently.

From research platform to clinical value

At Xijing Hospital, the platform supported research that predicts driver mutations in clear-cell renal cell carcinoma from routine HE slides, including VHL, PBRM1, BAP1 and SETD2. The AUC values reached 0.87, 0.84, 0.95 and 0.88, with BAP1 and SETD2 exceeding a general pathology foundation model baseline.

At Shanghai Pulmonary Hospital, the platform supported frozen small-nodule subtype classification and STAS recognition. The models were trained in two months, with AUC above 0.9 in both tasks, showing potential for improving intraoperative reading consistency and extending high-quality pathology capability to broader settings.

The hospital has also used the platform to develop and launch AI-assisted models for pleural and ascitic fluid cytology, bronchial brushing cytology, lung puncture cytology and EBUS-TBNA cytology. Around the frozen small-nodule program, more than 1,200 high-quality annotations and over 800 training iterations supported preliminary clinical validation.

A positive loop for hospital AI production

PathEngine has been deployed in nearly 30 backbone hospitals, including Xijing Hospital, Shanghai Pulmonary Hospital and Changzhou First People's Hospital. The broader collaboration network includes Shanghai Shenkang, Ruijin Hospital, Huashan Hospital, Fudan University Shanghai Cancer Center and others.

The larger signal is that every hospital's clinical practice can become part of the model evolution process. More real scenarios produce better data, better data trains more specialty models, and those models return to clinical validation and research use.