Artificial intelligence and machine learning have been used to accomplish several goals in the medical field, from reducing the spread of COVID-19 to verifying patient diagnoses.
These technologies have also had an important impact on preventing and detecting skin cancer. Recently, an algorithm was created to identify melanoma with impressive accuracy. Melanoma is the deadliest form of skin cancer, killing thousands of individuals in the United States each year. With early detection, the disease is often treatable, which is why technologies that can detect them are so beneficial. However, an early diagnosis is often difficult, leaving many people vulnerable.
The Need for New Methods of Examining Skin Lesions
The field of dermatology faces a lack of manpower when it comes to detecting melanomas early. There about 12,000 practicing dermatologists in the United States. Each of these professionals would need to examine more than 27,000 patients each year to screen the population of Americans with pigmented lesions suspicious for malignant melanoma. To address this problem, scientists have spent years trying to develop computer systems that can analyze skin lesions and identify them as malignant or benign. However, these systems have largely proved ineffective.
Part of the problem is that computer systems have largely only been able to analyze a single lesion at a time, which is not how dermatologists work. Also, most individuals’ pigmented lesions are not identical—they may have varying patterns. Dermatologists analyze suspicious lesions in the context of others on a patient’s body to determine whether they could be malignant. This approach is often called the “ugly duckling” criteria. If a patient has a lesion that is outside the common pattern of other moles on their skin, then it is considered the ugly duckling and typically biopsied to determine if it needs to be removed and through what method. This decision depends on the nature of the neoplasm and its suspected depth.
How the New System Works to Judge Suspiciousness
A recently introduced system has applied the ugly duckling approach to lesion analysis. The system relies on a convolutional deep neural network to analyze lesions against others on the same patient. Researchers at the Wyss Institute for Biologically Inspired Engineering at Harvard University and the Massachusetts Institute of Technology developed the new technology, which has demonstrated an accuracy of about 90 percent in distinguishing suspicious from non-suspicious lesions using only photos. Plus, during clinical trials, a consensus was achieved 88 percent of the time with the dermatologists who were used as a control.
The system created by researchers was first constructed like others made for this same purpose. Using more than 33,000 images, machine learning taught the program how to distinguish suspicious lesions from benign ones. Once this was achieved, it was time to facilitate an ugly duckling analysis. To accomplish this, the researchers used extracted features to create a three-dimensional map of lesions in a particular image and calculate how far from “typical” the features of each were. Doing this makes it possible to quantify how different a given lesion is from others on a single person’s skin. This comparison opened the door for deep learning networks to learn how to scrutinize the difference among lesions on one body.
The Promise of the New Algorithm and the Path Forward
To test the new system, 135 photos from 68 patients were run through it. These same images were judged by three dermatologists, who assigned an “oddness” score to the individual lesions. The computer system agreed with the individual dermatologists about 86 percent of the time. This finding is important because dermatologists tend to agree with each other about the suspiciousness of a lesion around 90 percent of the time. Thus, the algorithm offers accuracy that is very nearly on par with that of trained dermatologists. The photos used in the study were all taken with a smartphone, which opens up the potential for patients to screen themselves with more accuracy and potentially find melanomas much earlier than they would otherwise.
The developers of the technology recognize that it could serve a lot of good and thus have already made the algorithm open-source on GitHub. This step could help develop the algorithm even further and make it ready for commercial use in the near future.
Yet, there are several hurdles still to overcome before the technology becomes commercially viable. For example, the algorithm needs to demonstrate efficacy across the full range of different skin colors before it becomes commercially viable. Soon, the team hopes to partner with medical centers across the nation to test the efficacy of the system. Once overall efficacy has been demonstrated, the team proposes that this tool could be used by primary care physicians to determine whether a biopsy or a referral to a dermatologist is warranted.