Self-Supervised AI Improves Diagnostic Accuracy for Melanoma with Low Pathologist Agreement

Study results on new artificial intelligence (AI) that predicts diagnostic agreement for melanoma highlight the potential of the technology to improve diagnostic accuracy for this deadliest form of skin cancer and other diseases with low pathologist concordance.

Proscia’s (Philadelphia, PA, USA) retrospective study “Using Whole Slide Image Representations from Self-Supervised Contrastive Learning for Melanoma Concordance Regression” demonstrated the AI’s performance on 1,412 whole slide images of skin biopsies. Each image was assessed by three to five dermatopathologists to establish a concordance rate. The R2 correlation between the technology’s predictions and the dermatopathologists’ concordance rates was 0.51. Proscia’s research also indicates that the same AI could be extended to other diagnoses that demonstrate low pathologist agreement. This includes breast cancer staging as well as Gleason grading of prostate cancer, which is used to evaluate the aggressiveness of the disease. Both often play an important role in informing treatment decisions.

Image: Study results on new artificial intelligence predicts diagnostic concordance for melanoma (Photo courtesy of Proscia)

In addition to this study, Proscia plans to conduct additional research illustrating the potential benefits of AI in helping pathologists to diagnose melanoma, including:

• Lowering the misdiagnosis rate for difficult cases. Melanoma often presents like benign mimickers, causing pathologists to disagree on its diagnosis 40% of the time. As cases are often evaluated by only one pathologist, AI that predicts concordance with multiple experts could help to improve diagnostic accuracy by serving as a second set of eyes.
• Accelerating turnaround times for critical results. Over 15 million skin biopsies are taken annually in the U.S., each of which may display one of hundreds of diagnoses. AI that predicts diagnostic agreement could flag cases that were likely to be challenging, driving efficiency gains by suggesting additional testing to provide a more complete look prior to pathologist review.
• Reducing costs and distress for patients. Frequent over-diagnosis of melanoma not only results in additional costs for health systems but also leads patients to pay for unnecessary treatment and cope with the stress of believing they have a life-threatening disease. Increased diagnostic accuracy could help to eliminate these burdens.

“With this study, we have laid the groundwork for a new use case of AI in pathology that could have a tremendous impact on patient outcomes,” said Sean Grullon, Proscia’s Lead AI Scientist and lead author of the study. “Our technology relies on self-supervised learning to recognize incredibly subtle patterns, demonstrating the power of one of the most advanced approaches in AI.”

Related Links:
Proscia