Artificial intelligence software can now detect signs of the deadliest type of skin cancer with 100 percent accuracy, potentially detecting some of the deadliest cases at earlier stages when they have a better chance of being treated.
Dermatologists unveiled the third generation of their AI system that was able to identify all 59 cases of melanoma, which typically appear as irregularly shaped moles of various colors.
And out of 190 cases of skin cancer, including melanoma and less fatal forms, the AI software missed only one.
Melanoma is the deadliest skin cancer and more than 57,000 people died worldwide in 2020 alone. However, non-melanoma cancers, including squamous cell carcinoma and basal cell carcinoma, which are far less deadly but far more common, cause more deaths worldwide.
Artificial intelligence has become increasingly controversial in healthcare due to privacy and ethics concerns, but medical researchers are seeing it as a potential tool to streamline administrative tasks and shorten time to diagnosis to give patients the best chance of survival.
The researchers behind the study had worked for three years to fine-tune the technology before achieving peak accuracy in melanoma and near-complete accuracy in other skin cancer diagnoses
Skin cancer is the most common cancer in the United States. It is estimated that one in five people will develop it during their lifetime
UK dermatologists have been working on fine-tuning AI cancer detection software for the past three years, “training” the IT using patient data from consultations with doctors and images of their cancers.
More than 22,350 people in the UK tested the AI’s ability to detect their cancers and pre-cancerous lesions over a period of two and a half years.
Dermatologists and medical photographers have loaded the system with patient data, including photos of cancers, to teach it what to look for.
With the latest version of the software, more than 1,000 patient consultations have been performed to improve diagnostic accuracy.
The AI then sorted the data to distinguish between noncancerous lesions and possible cancers or “malignancies.” Dermatologists then checked the AI software’s diagnoses.
It diagnosed all 59 cases of melanoma as well as 99.5 percent of all skin cancers overall, including melanoma and non-melanoma cancers, a significant improvement over the first version’s accuracy of 83.8 percent.
Additionally, precancerous growths on the skin were correctly identified nearly 93 percent of the time, a 38 percentage point increase in accuracy compared to the first version.
Dr. Kashini Andrew, lead author of the study and a consultant at University Hospitals Birmingham NHS Foundation Trust in the UK, said: “This study has shown how AI is rapidly improving and learning, with high accuracy directly attributable to improvements in AI training. “Techniques and the quality of data used to train AI.”
He added that the technology has the ability to “free up more time for patients who need urgent attention.”
When examining the skin for cancerous growths, dermatologists follow the ABCDEs: asymmetry, border, color, diameter and development.
Most melanomas, the rarest but most dangerous form of skin cancer and the most likely to spread, appear as asymmetrical (the “A” in the ABCDEs) shaped moles with uneven edges.
Irregular edges of a mole, the “B” in ABCDEs, can also indicate melanoma. The edges of a normal birthmark are even and smooth.
A suspicious mole often contains multiple shades of tan, black, or brown, as well as pink, red, or purple spots, and becomes more colorful as the cancer progresses.
Melanomas are typically slightly larger than a pea or a pencil eraser, about six millimeters in diameter.
Melanomas grow in two phases, horizontal and vertical. This represents the “E” in the ABCDEs of melanoma diagnosis.
The horizontal phase can last years before the mole becomes dangerous and invasive, spreading to lymph nodes and organs. However, in a later phase, the lesion grows vertically and develops into a tumor, which can spread to other parts of the body and is potentially fatal.
Non-melanoma skin cancer manifests itself differently and often begins as waxy or red bumps on the skin. The researchers said they considered all other types of skin cancer, including those listed below.
Merkel cell carcinoma, the type of skin cancer that took the life of Jimmy Buffett, is a rare and aggressive form of skin cancer. It grows quickly and often spreads to the lymph nodes before affecting other organs.
Cancer of the Merkel cells, which are found in the outermost layer of skin, the epidermis, is about 40 times rarer than melanoma. While approximately 98,000 Americans receive a melanoma diagnosis in a year, only 3,000 people receive an MCC diagnosis annually.
Because early action is critical, annual screening is recommended. In fact, 99 percent of patients who detect their melanoma early and begin treatment survive five years or longer after diagnosis.
Basal cell carcinoma, which typically occurs on sun-exposed areas of the body such as the hands, neck, arms, and legs, often presents as a waxy lump or a small, smooth, shiny, or pale growth.
However, it does not always appear raised and may resemble a flat scar.
Another type of non-melanoma, squamous cell carcinoma, typically appears as a red, scaly patch of skin that sometimes bleeds. It may also appear as a raised scar.
People with a history of heavy sun exposure are most likely to develop this type of skin cancer.
Of all the basal cell carcinoma cases diagnosed, the AI software only missed one. But a “safety net” system, in which a doctor checked what the technology had missed, was able to detect the cancer.
This led researchers to make a major caveat that AI capabilities are not yet advanced enough to replace a practicing doctor in accurately identifying problems.
Dr. Irshad Zaki, consultant dermatologist at University Hospitals Birmingham NHS Foundation Trust and co-author of the study, said: “We would like to emphasize that AI should not be used as a standalone tool for skin cancer detection and that AI is not a replacement for consultant dermatologists.
“AI is currently not an independent tool in dermatology. “Our data shows the great potential of AI for future healthcare.”
The researchers’ results were presented at the 2023 congress of the European Academy of Dermatology and Venereology (EADV) in Berlin.