In a study, the results of which were published July 15 in Nature Medicine, the deep learning model analyzed slides from more than 15,000 patients with prostate, skin and breast cancer. The AI was able to detect cancer with 100 percent sensitivity and exclude almost all cancer-negative slides, allowing pathologists to focus only on the most critical slides for malignancy characterization.
The model is believed to be the first to achieve clinical-grade levels of accuracy in pathology. In contrast to other AI models, MSK’s is “weakly supervised,” meaning it was trained on a vast set of only minorly annotated slides, rather than fewer slides with more extensive annotation; the researchers suggest that this method helped the model better account for variability between various forms of cancer.
More articles about AI:
6 updates in AI ethics: new guidelines, institutes & more
Beth Israel Deaconess’ open-source patient database enabled a decade of AI research
VA appoints 1st director of AI, plans to expand AI research: 4 notes
