Using Artificial Intelligence To Detect Melanoma

Melanoma is the deadliest form of skin cancer and can be challenging to diagnose. In order to improve the rate at which treatment can be accessed, detection and diagnosis rates must also improve. The volume of skin biopsies continues to rise amid a declining pathologist population, slowing the rate of diagnosis and therefore, treatment. Earlier this month, Proscia released study data in which artificial intelligence (AI) was used to detect melanoma with a high degree of accuracy.

To find out more about how AI can improve melanoma diagnosis, Technology Networks spoke to Dr. Kiran Motaparthi, director of dermatopathology and clinical associate professor of dermatology at the University of Florida and Julianna Ianni, PhD, vice president of AI R&D, Proscia.

Kate Robinson (KR): What difficulties are associated with melanoma diagnosis?

Kiran Motaparthi
(KM): Some cases of melanoma can be challenging to distinguish from nevi, which are common, harmless (benign) skin lesions. Given that melanoma is also one of the most potentially lethal skin cancers, missing a melanoma can be consequential to a patient.

KR: Can you explain what a Pathology Deep Learning System (PDLS) is and how it works?

Julianna Ianni
(JI): When we say “Pathology Deep Learning System,” we’re referring to the deep learning system that we validated in the study. This system is applied to pathology data, so we call it a Pathology Deep Learning System. This system is trained by exposing it to examples of whole slide images of pathology specimens. Each specimen is associated with its own ground truth label, which indicates the type of patterns present in the tissue on the slide(s). By exposing the system to many of these examples of specimens and labels, the system learns the patterns that are associated with each label.

KR: Is the process of sample preparation the same with the addition of a PDLS?

The process is the same. In our study, the PDLS ran just before the pathologist viewed a case and after it was prepared. What’s really different is how much the pathologist knows about a case before reviewing it. Without the PDLS, the pathologist reviews cases at random. The PDLS classifies the case based on diagnostic attributes. Keeping with our example, the PDLS might run and tell the pathologist that a case is likely to be melanoma. In doing so, it could aid the pathologist to prioritize review for that type of case.

I should note that the PDLS can only run on digitized images of slides. To use the PDLS, or any AI application, a laboratory must already be using digital pathology. Unlike traditional pathology, where samples are affixed to glass slides, digital pathology centers around whole slide images of these glass slides. There are many reasons why a laboratory might want to go digital, and the ability to leverage AI is a key one. Among the others, it’s easier to share and collaborate when using whole slide images than when using glass slides.

KR: Can you describe the process by which your PDLS will be used to improve melanoma diagnosis?

Our study findings indicate that the PDLS holds the promise to improve diagnostic accuracy and deliver faster results to patients, and we are conducting additional research to further explore this potential.

AI that automatically detects melanoma could serve as an adjunctive aid to the pathologist. This could help the pathologist to make a more accurate diagnosis and lead to better patient outcomes as a result. AI could also flag high-risk cases to the pathologist for earlier review or ensure that the right cases get sent straight to someone with subspecialty expertise. The pathologist could then prioritize patients with the most clinically impactful diagnoses so that they could begin treatment sooner, also helping to improve patient outcomes.

KR: Could a PDLS be used to assist in the diagnosis of other diseases, including those which aren’t cancerous?

Yes. In fact, the PDLS that we studied also classifies skin biopsies that aren’t cancerous. The specific diseases a PDLS could recognize would just depend on how it was trained and the data available.


Original post:

Leave a Reply

Your email address will not be published. Required fields are marked *