FAHZU Uses OmniPT in Cancer Pathology: AI Counts Slides, Pathologists Make Final Judgments

Zhejiang’s provincial government website reported on December 26, 2024, that the First Affiliated Hospital of Zhejiang University School of Medicine was using the AI pathology assistant OmniPT. The Hangzhou hospital’s work focuses on high-incidence cancers including gastric, colorectal and cervical cancers. The system combines vision and language models to support pathologists through human-computer interaction.

From specimens to cancer classification and grading

When tissue or cell samples reach the laboratory, pathologists assess pathological changes and the nature of the disease. OmniPT assists slide evaluation, accelerating specimen analysis and helping identify details that a fatigued doctor might overlook, according to the report.

The hospital’s laboratory tests covered cancer classification, grading, vascular and neural invasion, and markers indicating the disease’s future course. The hospital reported 80–90% accuracy for analyses and predictions across tested cancer types. This figure describes the reported laboratory testing, rather than diagnostic accuracy or treatment outcomes for every patient.

Mitosis counting: 30–60 minutes manually, under 10 seconds with AI

Pathology head Zhang Jing used mitosis counting in glioma diagnosis as an example. Gliomas can occur in the brain or spinal cord, and counting these features on slides is part of diagnostic work. He said manual counting on one slide under a high-power microscope takes at least 30 minutes to an hour, while OmniPT can complete that task in under 10 seconds.

The comparison concerns slide counting, not the total time from specimen submission to a completed pathology report. Uncertain features are escalated to a pathologist. Zhang illustrated this with 10 uncertain mitotic figures referred to him for a final decision, rather than resolved by the system alone.

Repetitive work assisted, diagnosis led by doctors

The report says OmniPT handles more than 90% of repetitive tasks, leaving features requiring further assessment to pathologists. Development centered on clinical needs, faster slide evaluation and repeated work.

Zhang described an assistant directed by doctors, who retain final judgment. The implemented workflow combines computation, escalation of uncertain images and medical review; it does not treat a machine analysis alone as the patient’s final diagnosis.