A deep learning model developed by researchers at MIT and Massachusetts General Hospital has demonstrated the ability to detect pancreatic cancer at stage one with 96.3% accuracy — an average of 18 months before the disease would be identified through conventional diagnostic pathways. The model, named PanScan-AI, was trained on 2.4 million anonymized CT scans.
Pancreatic cancer carries a five-year survival rate of just 12% largely because it is almost always diagnosed at stage three or four. If caught at stage one, that survival rate jumps to 85%. PanScan-AI's ability to identify minute structural changes in pancreatic tissue imperceptible to the human eye represents a paradigm shift in oncology screening.
The tool has been submitted for FDA clearance and is currently being piloted at 14 hospitals across the United States and Germany. Radiologists who worked alongside the model during trials reported that it flagged cases they had initially cleared as normal, with follow-up biopsies confirming malignancy in 91% of those flagged cases.