A deep learning artificial intelligence system developed at Memorial Sloan Kettering Cancer Center has achieved 97.3% diagnostic accuracy across 47 rare cancer types — outperforming the average board-certified oncologist by a margin of 22 percentage points in blinded comparative trials, according to a study published in NEJM AI.
How the System Works
The AI, trained on 2.3 million histopathology slides and 890,000 radiology images annotated by expert oncologists across a 15-year period, analyzes microscopic tissue patterns invisible to the human eye. It can detect molecular subtypes of cancer from standard histology slides that would normally require expensive and time-consuming genetic sequencing.
Performance in Rare Cancers
The most dramatic improvements were seen in rare cancer subtypes, where even specialized oncologists struggle due to limited case exposure. In cholangiocarcinoma (bile duct cancer), the AI achieved 98.1% accuracy versus 64% for specialists. In peritoneal mesothelioma, 96.8% versus 71%.
"The AI has effectively seen more rare cancer cases than any human oncologist alive. That experience advantage is what we're seeing translated into diagnostic accuracy." — Dr. James Park, Lead Researcher
Integration Into Clinical Practice
The system is designed to augment, not replace, oncologists. It generates a ranked differential diagnosis with probability scores and highlighted tissue regions of concern, which the oncologist reviews and confirms. Early pilots in 8 major cancer centers show an average 11-day reduction in time-to-diagnosis for complex cases.
Regulatory Status
The FDA granted Breakthrough Device designation in September 2025. Full approval is expected in mid-2026, with rollout to 1,200 cancer centers planned through 2027.