Pathology AI analyzes entire slides, outperforming existing models by 19.1 percentage points

Researchers in South Korea have developed a generative pathology artificial intelligence (AI) system that analyzes whole slides to help diagnose cancer. It showed substantially higher accuracy than existing models.

A research team led by Professor He Junjun of Shanghai Jiao Tong University and Professor Yu Rongshan of Xiamen University reported the findings on the 1st (local time) in the international journal Nature Cancer.

The AI system, called SlideChat, was designed to analyze whole-slide images across cancer types. It combines patch-level and slide-level pathology encoders with a pretrained large language model (LLM).

Existing pathology AI systems have been limited to interpreting only parts of a slide, making it difficult to capture the full tissue context needed for diagnosis. SlideChat was trained on 274,233 multimodal instruction samples to learn the relationship between whole slides and diagnostic reports.

Performance was evaluated across 31 cancer types and five independent cohorts. On 8,836 multiple-choice questions, SlideChat was 19.1 percentage points more accurate than the top-performing existing model.

For report generation, it scored 7.7 percentage points higher on the translation evaluation metric METEO, and it received the highest expert scores across all five evaluation dimensions on 129 open-ended questions. The researchers said SlideChat has the potential to support pathology diagnosis, medical education and clinical decision-making.